Let’s be real for a second. We’ve all been there—staring at a blank slide, cursor blinking, deadline creeping closer, and somehow the presentation still looks like it was made in 2014. You’ve got the ideas. You’ve got the data. But turning all of that into a deck that actually looks like your brand? That’s where things fall apart.

Well, that pain point just got solved. Pitch—the AI presentation workspace already trusted by millions of teams worldwide—just launched Pitch Agent, and it’s not just another “add some AI sparkle” feature. This is a full-blown conversational AI assistant that lives right inside your presentation editor, and it’s changing the entire game for how we create, refine, and deliver slide decks.

Get Pitch Now!

😩 The Problem: AI Presentations That All Look the Same

If you’ve tried any of the AI presentation tools floating around, you probably noticed the same issue: they’re fast, sure, but the output is generic. They slap your logo on a cookie-cutter template, swap in your brand colors, and call it “on-brand.” But every deck ends up looking interchangeable—the same rigid rows, the same columns, the same lifeless layouts.

That’s not on-brand. That’s re-theming. And there’s a massive difference.

Your brand has patterns—unique visual DNA that your designers spent time crafting. The way images interact with text, the specific layouts that feel likeyou. Generic AI tools don’t understand that. Pitch Agent does.

Get Pitch Now!

🤖 So What Exactly Is Pitch Agent?

Pitch Agent is a conversational AI assistant built directly into Pitch that can do three core things:

• Generate on-brand presentations from a prompt—you give it a brief (and optionally attach files like PDFs, CSVs, or text docs for context), and it creates a fully designed, multi-slide deck using your actual brand template. Not a generic layout. Your template, with your custom layouts.

• Edit and refine slides through conversation—want to swap an image? Split a dense slide into three? Merge a few slides into a summary? Rewrite a section? Just ask. It understands context and makes the changes while keeping everything visually consistent.

• Chat with your deck—and this is where it gets really interesting. Pitch Agent has full awareness of your entire presentation, so you can ask it questions like “What are our key claims?”, “Does the narrative flow for this audience?”, or “Write me a follow-up recap email based on this deck.” It becomes a pitching partner, not just a slide generator.

🎯 What Makes Pitch Agent Different from Everything Else?

I’ve tested a lot of AI presentation tools. Here’s what stood out about Pitch Agent:

1. It Builds from Real Templates, Not Generic Grids

Most AI tools work with a fixed set of layouts — title slide, two-column, three-column, image-left, done. Pitch Agent actually reads your chosen template and learns from its fully custom layouts. The result? Slides that look like your design team made them, not like an AI spit them out. It even draws from a library of 150+ additional layouts to fill in any gaps your template might have.

2. AI-Generated Images That Match Your Visual Style

Here’s something that genuinely surprised me. When Pitch Agent generates a presentation, it also produces images that match the photography and art direction of your template. So if your brand template uses moody, dark-toned editorial photography, the AI-generated visuals follow that same aesthetic. No more awkward stock photos that clash with your deck’s vibe.

3. Everything Is Fully Editable

Unlike some AI tools that lock you into their output, everything Pitch Agent creates is completely editable inside Pitch’s powerful slide editor. Adjust layouts freely, tweak text, move elements around—you maintain total control. The AI gives you a strong starting point, and you polish it to perfection.

4. It’s Not Just Generation — It’s the Entire Lifecycle

This is the crucial distinction. Pitch isn’t just an AI slide maker. It’s a complete presentation workspace with brand and asset management, real-time collaboration, advanced sharing options (including digital sales rooms!), and rich engagement analytics. Pitch Agent slots into this ecosystem beautifully—you generate a deck, collaborate with your team, share it via a live link, and track who viewed which slides and for how long.

Get Pitch Now!

💡 Real-World Use Cases That Got Me Excited

Here’s where the rubber meets the road. These are the scenarios where Pitch Agent truly shines:

Create a Full Deck from a Single Prompt

Select a template, type something like “Create a renewal-focused QBR for Omnexis including value delivered, adoption progress, and a 90-day plan to expand usage,” attach your account overview PDF and usage report CSV, and watch Pitch Agent build an entire editable deck. The time savings versus starting from a blank page are enormous.

Personalize a Deck for a Specific Client

Got a generic sales deck that needs to feel tailored for a prospect? Feed Pitch Agent your call notes, the company’s details, or the role you’re pitching to, and it makes targeted edits while keeping the deck’s overall structure and narrative intact. Faster, more relevant decks with way less manual rework.

Turn Existing Documents into Slides

Have a long PDF report, a text document, or a CSV full of data? Bring that material into Pitch as context, and let the AI repurpose it into slides instead of you rewriting everything from scratch. No copy-paste errors. No reformatting headaches.

Generate Follow-Up Content from Your Deck

This one is a game changer for sales teams. After a meeting, you can ask Pitch Agent to draft a post-meeting recap email, a summary of key talking points, or a personalized follow-up—all pulled directly from the content in your slides. Your messaging stays consistent from the deck to the inbox.

Pressure-Test Your Story Before You Present

Ask questions like “What objections are we answering here?”, “What’s the 30-second elevator pitch of this deck?”, or “Does this narrative flow make sense for a CFO audience?”Pitch Agent gives you instant, context-aware answers. It’s like having a presentation coach built right into your tool.

Get Pitch Now!

⚡ Getting Started is Ridiculously Easy

There are three ways to jump into Pitch Agent:

• New presentation prompt box—when you create a new presentation, you’ll see a prompt box right away. Select a template, write your brief, attach files if you want, and hit Generate. You’ll get a full deck ready to edit.

• Assistant sidebar—inside the editor, click the Pitch Agent button in the top right corner to open a conversational sidebar. Go back and forth with the AI to create slides, make edits, answer questions, or draft follow-up content.

• Brand template generator—set up a branded template in seconds. Just enter your company’s domain, and the AI pulls in your brand cues automatically. Confirm your colors, logo, and fonts, then start generating on-brand decks instantly.

And the best part? Pitch is a freemium product. You can create presentations for free. Pitch Agent is available on all plans — including the free tier — with different AI credit allocations depending on your plan. Paid plans (Plus, Team, Business, and Enterprise) unlock more features and more credits.

📂 What Can You Feed It?

At launch, Pitch Agent supports the following attachment types for context: PDFs (with full text extraction), Markdown files, CSV files (great for tables and charts), and plain text files. Just drag and drop your source materials, and the AI will use them to generate more targeted, data-rich presentations.

✨ Bonus: Quick Actions for Instant Polish

Pitch Agent works alongside Pitch’s existing quick actions—single-purpose AI tools for things like removing image backgrounds, shortening text, rewriting slide copy to match your brand tone, turning text into charts, and expanding images. Think of quick actions as your finishing touches, while Pitch Agent handles the heavy lifting.

🏁 The Bottom Line

I’ve seen a lot of AI tools come and go in the presentation space. Most of them optimize for speed and sacrifice quality. Pitch Agent is different because it doesn’t just generate slides—it generates your slides, in your brand, with your templates, and then lets you refine them through natural conversation. And because it’s built into Pitch’s full presentation workspace, the work doesn’t end at generation. You collaborate, share, present, track engagement, and iterate — all in one place.

Whether you’re a startup founder crafting pitch decks, a sales team personalizing proposals, an agency delivering client-facing presentations, or a marketer launching campaigns — Pitch Agent is going to save you serious time while making your output look significantly better.

Get Pitch Now!

That’s it for today’s spotlight, folks. If you found this useful, share it with a colleague who’s still fighting with PowerPoint at 2 AM. 😄

Hey there, AI enthusiast! 👋

Welcome back toAI Buzz!— This might be the most consequential week in AI history. The Pope just issued a 42,000-word encyclical on artificial intelligence. Anthropic is closing a $30 billion round that would make it the world’s most valuable AI startup. OpenAI filed its IPO paperwork. And the President scrapped an AI safety executive order after tech CEOs called him directly. Oh, and the most beloved AI researcher alive just switched teams. Let’s dive in.

Distill AI

500+ AI papers drop daily. You don't have time for all of them. Distill AI picks the ones that matter, delivers to your inbox every morning ☕ — and lets you chat with any paper for instant answers.

🔥 TODAY’S TOP STORIES

Anthropic Closes $30B Round at $900B+ Valuation — Now the World’s Most Valuable AI Startup

Why it matters: Anthropic is closing a $30 billion funding round at a valuation above $900 billion, co-led by Sequoia, Dragoneer, Greenoaks, and Altimeter — each investing roughly $2 billion. This vaults Anthropic past OpenAI’s $852 billion March valuation for the first time ever. The company projected $10.9 billion in Q2 revenue and its first quarterly operating profit, up 130% from Q1. Annualized revenue is now approaching $45 billion — a fivefold increase from $9 billion at end of 2024.

💡 The takeaway: This is no longer a hype valuation — it’s backed by real revenue and profitability. Anthropic just proved that a safety-focused AI lab can also be a financial juggernaut.

Pope Leo XIV Issues First-Ever Papal Encyclical on AI

Why it matters: Pope Leo XIV today released “Magnifica Humanitas,” a sweeping 42,000-word encyclical on AI and human dignity — co-presented with Anthropic co-founder Chris Olah at the Vatican. Signed on the 135th anniversary of Rerum Novarum (the landmark teaching on workers’ rights during the Industrial Revolution), the document calls AI “the Industrial Revolution of our time” and urges governments and corporations to slow development and ensure AI serves humanity, not the powerful few.

💡 The takeaway: The Vatican choosing Anthropic — not Google, not OpenAI — as its partner is a deliberate signal about which lab it considers credible on safety. This will shape public discourse on AI ethics for years.

OpenAI Files Confidential IPO Paperwork, Targeting $1 Trillion Valuation

Why it matters: OpenAI confidentially filed its S-1 IPO prospectus with the SEC on May 22, with Goldman Sachs and Morgan Stanley leading the deal. The company is targeting a public listing as early as September 2026 at a valuation between $852 billion and $1 trillion. This comes as both OpenAI and Anthropic are expected to go public this fall — potentially the two largest AI IPOs in history, within months of each other.

💡 The takeaway: Two near-trillion-dollar AI companies going public simultaneously will force unprecedented financial transparency on an industry that’s largely avoided it. Watch for the public S-1 — that’s when the real numbers drop.

Distill AI

When drones learn to listen

⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐…

Read more

9 days ago · Distill AI

Trump Scraps AI Safety Executive Order After Tech CEO Phone Calls

Why it matters: President Trump abruptly canceled the signing of a landmark AI executive order, telling reporters it could “get in the way” of America’s AI lead over China. The order would have created a voluntary pre-release vetting process for advanced AI models — a direct response to cybersecurity concerns triggered by Anthropic’s Claude Mythos. Reports indicate Mark Zuckerberg, Elon Musk, and David Sacks called the president directly in the hours before the scheduled signing.

💡 The takeaway: No federal AI framework means states and the EU AI Act fill the vacuum. This is short-term bullish for AI labs (faster release cycles) but long-term uncertain (fragmented regulation).

NextEra Buys Dominion Energy for $67B — Largest Utility Merger in U.S. History

Why it matters: NextEra Energy announced a $67 billion all-stock deal to acquire Dominion Energy, creating the world’s largest regulated electric utility. The deal is explicitly driven by AI power demand — data centers are projected to consume 15–25% of US electricity by 2030. Dominion powers Northern Virginia’s “Data Center Alley,” the world’s largest data center market. The combined enterprise value of $420 billion makes it the third-largest U.S. energy company behind Exxon and Chevron.

💡 The takeaway: Power availability — not model capability — is now the primary bottleneck for AI scaling. The AI boom is reshaping industries far beyond tech.

Andrej Karpathy Joins Anthropic — The Biggest AI Talent Move of 2026

Why it matters: Andrej Karpathy — OpenAI co-founder, former Tesla Autopilot lead, and the most beloved AI educator alive — announced he’s joining Anthropic’s pre-training team. He’ll build a new team focused on using Claude to accelerate Claude’s own training — a form of recursive self-improvement. He’s the latest in a pattern: six CTOs from billion-dollar companies have taken individual contributor roles at Anthropic in the past year.

💡 The takeaway: When the most respected researcher in the field chooses Anthropic over every other option, that’s a signal every senior AI researcher will notice. Talent follows conviction.

Distill AI

What 6 chatbots get wrong about the news

⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐…

Read more

9 days ago · Distill AI

⚡ QUICK HITS

• Google Search’s Biggest Update in 25 Years: Google rebuilt its search box with AI, now generating custom visuals, interactive graphics, and mini-apps directly on the search page. The 10-blue-links era is functionally over.

• Chinese Models Hit 60% on OpenRouter: Kimi K2.6, DeepSeek V4, GLM-5.1, and Qwen 3 now account for 60% of all AI usage on OpenRouter. The open-weights tier is now Chinese-led.

• Adobe, Canva & CapCut Integrate Gemini: Users can now generate AI content inside Gemini and refine it directly in professional editing tools. Canva is live for Gemini AI Ultra subscribers.

• Intuit Lays Off 17% as AI Restructuring Spreads: Following Salesforce, Microsoft, Google, and Meta’s 8,000 cuts — Intuit is the latest enterprise software giant restructuring around AI.

• Anthropic Eyes Microsoft’s Maia 200 Chip: Anthropic is in talks to adopt Microsoft’s custom AI chip, making it the fifth silicon partner alongside NVIDIA, AWS Trainium, Google TPUs, and SpaceX compute.

• Anthropic’s $45B SpaceX Compute Deal Revealed: SpaceX’s IPO prospectus disclosed Anthropic is paying $1.25 billion per month through May 2029 for GPU compute — a $45 billion total contract.

📊 AI NUMBER OF THE WEEK

$900,000,000,000

👀 WHAT WE’RE WATCHING

The IPO race is on. Both OpenAI and Anthropic are expected to go public between Labor Day and Thanksgiving 2026. SpaceX has already filed. These could be three of the largest IPOs in technology history, all within months of each other. The AI industry is about to face a level of financial scrutiny it’s never experienced. We’ll be covering every filing, every number, every implication.

Hey there, AI enthusiast! 👋

Welcome back to AI Buzz! — your no-fluff, all-signal briefing on the AI stories that actually matter. This week, the AI arms race escalated to a whole new level: Anthropic is eyeing a jaw-dropping $900 billion valuation, Meta just dropped its first closed-source model, and Big Tech committed three-quarters of a TRILLION dollars to AI infrastructure. Buckle up — here’s everything you need to know.

Read time: 5 minutes

Anthropic Eyes $900B Valuation in Massive $50B Funding Round

Why it matters: If completed, Anthropic would surpass OpenAI’s $852 billion valuation to become the world’s most valuable AI startup. The round reflects explosive enterprise demand for Claude, with annualized revenue approaching $45 billion — a fivefold increase from $9 billion at the end of 2024. Products like Claude Code and Cowork are driving the surge. CEO Dario Amodei warned that SaaS firms failing to adopt AI could face bankruptcy, and the company is reportedly eyeing an IPO as early as October 2026.

💡 The takeaway: Enterprise AI adoption is no longer optional. A near-trillion-dollar private valuation signals that the market sees AI infrastructure as the next computing paradigm — not a bubble, but a foundation.

Meta Launches Muse Spark — Its First Closed-Source Flagship Model

Why it matters: This is a seismic strategic shift. Meta, the company that championed open-source AI with its Llama models (1.2 billion downloads), just went proprietary. Built under Chief AI Officer Alexandr Wang’s newly formed Superintelligence Labs, Muse Spark is competitive with GPT-5.4 and Claude Sonnet 4.6 on multimodal reasoning and agentic tasks. Meta also confirmed a staggering $115–$145 billion in AI capital expenditures for 2026 — nearly double last year.

💡 The takeaway: Meta is signaling that the open-source-everything approach has limits at the frontier. If you’ve built on Llama, it’s time to diversify. Watch how the developer community reacts.

Big Tech Commits $725 Billion to AI Infrastructure While Slashing Jobs

Why it matters: Meta, Amazon, Microsoft, and Alphabet have collectively committed roughly $725 billion in capital expenditures for 2026 — a 77% increase year-over-year — almost entirely earmarked for data centers, custom chips, and AI models. But the twist: they’re simultaneously cutting tens of thousands of jobs. Meta plans to cut 8,000 roles, Amazon has trimmed ~30,000, and Microsoft offered voluntary buyouts to 125,000 employees. Rising memory chip prices alone pushed budgets $55 billion higher than initial estimates.

💡 The takeaway: The era of “more people = more output” is ending. These companies are betting that compute, not headcount, is the future. Start thinking about how AI augments YOUR role.

OpenAI Crosses $25B in Revenue & Launches GPT-5.5 Instant

Why it matters: OpenAI surpassed $25 billion in annualized revenue and is exploring an IPO as early as late 2026. Meanwhile, the company launched GPT-5.5 Instant as the new default ChatGPT model, boasting 52.5% fewer hallucinated claims than its predecessor in high-stakes domains like medicine, law, and finance. The model is also more concise, uses fewer emojis, and can now search your past conversations, files, and connected Gmail for personalized answers. OpenAI is also testing ads in ChatGPT for the first time.

💡 The takeaway: Ads in ChatGPT could change the user experience. Keep an eye on whether this shifts OpenAI from a product company to an ad-supported platform.

Novo Nordisk Partners with OpenAI to Supercharge Drug Discovery

Why it matters: The Danish pharma giant announced a strategic partnership with OpenAI to integrate AI across its entire business — from drug discovery and clinical trials to manufacturing and supply chains. Full deployment is planned by end of 2026. The goal? Accelerate new obesity and diabetes treatments as Novo fights to regain ground against Eli Lilly. CEO Mike Doustdar emphasized the aim is to “supercharge” scientists, not replace them. OpenAI will also upskill Novo’s global workforce.

💡 The takeaway: This is AI’s biggest pharma integration to date. If successful, it could cut drug development timelines from 10 years to 5 — a game-changer for patients worldwide.

Five Eyes Agencies Release Joint Agentic AI Security Guidance

Why it matters: Cybersecurity agencies from the US, UK, Australia, Canada, and New Zealand jointly released a 30-page guidance document titled “Careful Adoption of Agentic AI Services.” It identifies five risk categories for AI agents in critical infrastructure: privilege, design/configuration, behavior, structural, and accountability. The guidance stresses incremental deployment, strict access controls, and rigorous human oversight — warning that AI agents are already operating inside critical infrastructure with dangerously little governance.

💡 The takeaway:As AI agents get more autonomous, governments are catching up on security frameworks. If your company uses AI agents, this document is essential reading.

⚡ QUICK HITS

• China’s Coding Sprint: Four Chinese labs released open-weights coding models within 12 days — GLM-5.1, MiniMax M2.7, Kimi K2.6, and DeepSeek V4 — all at roughly the same capability ceiling at dramatically lower inference cost than Western competitors.

• Eli Lilly’s AI Supercomputer: Lilly inaugurated “LillyPod,” pharma’s most powerful AI supercomputer with 1,016 Blackwell Ultra GPUs delivering 9,000+ petaflops. Goal: simulate billions of molecular hypotheses in parallel to cut drug timelines in half.

• Apple’s $250M Siri Settlement: Apple agreed to pay $250 million over delayed Siri AI features. Claims open within 45 days of May 5 for qualifying device owners.

• Global AI Usage Hits 17.8%: Microsoft reports AI usage rose to 17.8% of the world’s working-age population in Q1 2026, with 26 economies exceeding 30%. The UAE leads globally at 70.1%.

• Cohere + Aleph Alpha Merger: Cohere announced a merger with Germany’s Aleph Alpha, positioning the combined entity as a sovereign AI alternative to the US-China duopoly.

• JPMorgan Goes All-In: JPMorgan Chase reclassified AI from “experimental R&D” to core infrastructure, dedicating a $19.8 billion tech budget and 2,000 staff to AI development.

📊 AI NUMBER OF THE WEEK

$725,000,000,000

That’s how much Big Tech has committed to AI infrastructure spending in 2026 alone. To put that in perspective, it’s bigger than the GDP of Switzerland or Turkey. The message is clear: the AI infrastructure buildout is the biggest capital deployment since the Internet itself.

👀 WHAT WE’RE WATCHING

The AI pricing war is heating up. Five major pricing changes from three of the four biggest AI providers all landed in April. The original design of subscription plans is being challenged by evolving capabilities and usage patterns. As inference costs plummet and models get commoditized, expect dramatic shifts in how AI products are packaged and sold. We’ll be tracking this closely in upcoming editions.

Part III of a three-part series on AI agent security fundamentals.

Summary

Security researchers have already demonstrated self-replicating AI worms, persistent compromise mechanisms, and multi-stage attack chains against AI agents. These vulnerabilities have been tested against a range of AI systems, including Claude Code, ChatGPT and Gemini. Understanding these emerging threats is essential for anyone using, building or deploying agentic AI systems.

Want to go deeper? Subscribe to the Agentic AI Briefing, a weekly publication covering emerging trends, threats, defensive strategies, and real-world lessons from the frontier. Sign up here: https://aisecurityguard.io/agentic-ai-briefing

Subscribe to the Agentic AI Briefing

Beyond Simple Prompt Injection

Many people may have heard of prompt injection. These are attacks that trick an AI into ignoring its instructions or executing malicious ones. Protecting against prompt injection is important, but there is a larger threat landscape facing agentic systems.

Researchers Ben Nassi, Bruce Schneier, and Oleg Brodt recently mapped real-world AI security incidents into a framework they call the Promptware Kill Chain. It reveals that prompt injection is just stage one of a seven-stage attack mechanism.

The full chain:

• Stage 1: Initial Access: Malicious prompt enters context (direct or via documents)

• Stage 2: Privilege Escalation: Attacker bypasses safety guardrails (jailbreaking)

• Stage 3: Reconnaissance: Agent reveals its tools, permissions, and connected services

• Stage 4: Persistence: Attack embeds in memory or poisons retrieval systems

• Stage 5: Command & Control: Agent fetches updated instructions from attacker

• Stage 6: Lateral Movement: Attack spreads to other users, agents, or systems

• Stage 7: Actions on Objective: Data theft, fraud, or physical-world impact

The key insight: each stage enables the next. An attacker that achieves only initial access has limited impact. One that is able to establish a persistent connection to the agent’s device, and can command the agent to engage in tasks has access to many system resources, depending on the agent’s permissions.

Subscribe to the Agentic AI Briefing

Demonstrated Attacks

Researchers have demonstrated attacks against production systems that follow all or parts of the kill chain.

Morris II: The First AI Worm (2024)

Researchers created a self-replicating worm targeting email assistants using internal databases to augment their capabilities:

1. Attacker sends email containing adversarial prompt

2. Email gets stored in the agent’s retrieval database

3. When the victim asks about emails, the malicious prompt gets retrieved

4. The prompt jailbreaks the agent and exfiltrates data from other emails

5. The agent automatically replies to contacts, spreading the payload

6. Zero user interaction required after the initial email

Subscribe to the Agentic AI Briefing

“Invitation Is All You Need” (2025)

Researchers demonstrated attacks through calendar invites:

1. Attacker sends calendar invitation with embedded prompt injection

2. User asks their AI assistant “What’s on my calendar today?”

3. The prompt injection activates and compromises the assistant

4. The attack persists in the user’s workspace memory

5. Researchers demonstrated: location identification and video recording

Google acknowledged and patched this vulnerability after disclosure.

Why Sandboxing Alone Isn’t Enough

One way to reduce agent exposure to attacks is to sandbox it. However, in some cases this can be insufficient.

What Sandboxing Prevents:

• Agent exceeds filesystem permissions

• Agent executes arbitrary code

• Agent accesses forbidden network resources

What It Doesn’t Prevent:

• Agent is allowed to read files, is tricked into reading and transmitting details from sensitive ones

• Agent uses permitted tools in unintended ways

• Agent sends data through permitted channels (email, API calls)

The attack surface is the agent’s legitimate capabilities. If an agent can send emails and read documents, an attacker can trick it into emailing sensitive information. The sandbox sees only permitted actions. The intent is what’s malicious.

Subscribe to the Agentic AI Briefing

Why Pattern Matching Leaves Gaps

Another defensive strategy is creating a set of rules limiting what an agent can do, such as accessing certain IP addresses or file systems.

But there is a limitation to this approach. According to the authors of the Promptware Kill Chain research:

“Guardrails operate at the application layer, not the architectural layer. They function as pattern-matching defenses against known attack signatures rather than as enforcement of a fundamental boundary between instructions and data.”

An attack might bypass defenses because no signature or detection rule exists yet.

The asymmetry problem:

• Defenders must anticipate and block all possible injection techniques

• Attackers need to discover just one that works

Pattern matching raises the bar. But this method cannot detect every type of attack.

Supply Chain Risks

Beyond the kill chain, there’s a parallel threat: supply chain poisoning.

AI agents rely on external packages, plugins, and tool definitions. Malicious code in a dependency can inject attack capabilities before any prompt injection occurs.

Recent examples include typosquatted npm packages that targeted AI development tools and compromised MCP (Model Context Protocol) server definitions that included hidden instructions.

The Defense-in-Depth Imperative

The Promptware Kill Chain researchers’ conclusion:

“Assuming initial access will occur, practitioners must focus on limiting privilege escalation, preventing persistence, constraining lateral movement, and minimizing the impact of actions on the objective.”

This represents a fundamental shift. Instead of “prevent all prompt injection” (impossible), the goal becomes limiting damage at each subsequent stage.

Effective defense requires multiple controls:

• Input scanning to detect injection patterns before they reach the agent

• Output monitoring to flag data exfiltration attempts

• Memory integrity checks to detect poisoned context

• Network monitoring to block callbacks to attacker-controlled URLs

• Tool integrity verification to ensure plugins haven’t been compromised

• Behavioral anomaly detection to flag unusual action sequences

No single control covers all stages. The attacker needs one path through. The defender needs coverage across all of them.

What’s Next?

This three-part series covered the fundamentals: what AI agents are, the shadow AI problem already affecting most organizations, and the emerging threat landscape that security researchers are documenting.

But understanding threats is just the beginning. The next step is taking action.

The AI Security Action Pack provides 15 practical guides—each with concrete mitigation steps and installable skills your agents can use to protect themselves. It’s designed for humans who need to understand the risks and agents that need to operationalize the defenses.

Download the AI Security Action Pack: https://aisecurityguard.io/action-pack

This is the third installment of a three-part series on AI agent security fundamentals.

• Part I: What AI Agents Actually Are (And Why Security Teams Should Care)

• Part II: Shadow AI — The Security Risk Hiding in Plain Sight

• Part III: Understanding Emerging AI Security Threats — From Prompt Injection to AI Worms

This is the third installment of a three-part series on AI agent security fundamentals developed by Fard Johnmar, international best-selling technology and innovation author, builder and founder of Enspektos, LLC, a firm developing products and services for the autonomous AI future.

Part II of a three-part series on AI agent security fundamentals.

Summary

In April 2026, Vercel, a major cloud platform, was breached after an employee granted a third-party AI tool access to their Google Workspace. The attacker pivoted through the AI tool to hijack internal systems and demanded $2 million in ransom.

This isn’t an isolated incident. More than 80% of employees are already using unauthorized AI tools at work, and many have shared sensitive company data with these tools without approval. This “shadow AI” problem represents one of the largest unaddressed security risks in most organizations — and traditional security controls can’t catch it.

Want to go deeper? The AI Security Action Pack includes 15 practical guides covering the most common AI agent security risks, plus installable skills your agents can use to protect themselves. Each guide maps to the OWASP Top 10 for Agentic Applications.

Download the AI Security Action Pack: https://aisecurityguard.io/action-pack

Employees Are Already Using AI Agents

Here’s the uncomfortable reality: your colleagues have almost certainly adopted AI tools you don’t know about.

The numbers are striking. According to UpGuard’s 2025 “State of Shadow AI” report:

• More than 80% of workers use unapproved AI tools in their jobs

• 68% of security leaders — including CISOs — admit to using unauthorized AI in their daily workflows

• Senior leadership is 50% more likely to use shadow AI than junior employees

• Increased security training doesn’t reduce usage — it actually correlates with higher shadow AI adoption

And now we’re seeing what happens when one of those tools gets compromised.

Why It’s Happening

Shadow AI isn’t malicious. It’s pragmatic.

Employees face productivity pressure. AI tools deliver immediate results. The gap between “what IT provides” and “what helps me do my job” gets filled by whatever’s available.

Consider the typical scenario: An employee needs to summarize a complex document for a meeting in 30 minutes. The company hasn’t approved an AI tool for this. But ChatGPT is right there, free, and fast. What do they do?

According to BlackFog’s January 2026 research, 60% of employees believe using unsanctioned AI tools is worth the security risks if it helps them meet deadlines. And 63% think it’s acceptable to use AI without IT oversight if no company-approved option exists.

The Vercel employee who connected Context.ai to their Google Workspace probably had similar reasoning: it’s just a productivity tool. But that single decision created an attack surface that compromised an entire enterprise.

The Data Exposure Problem

Sensitive data is being exposed via AI systems. Employees have shared enterprise research or datasets with AI tools. Some have revealed employee data — names, payroll, performance information. Others input company financial statements or sales data into these systems.

Harmonic Security’s analysis of 22 million enterprise AI prompts in 2025 provides additional detail. Of the prompts containing sensitive data, the top categories were:

• Source code: 27%

• Legal content: 22%

• M&A data: 13%

• Financial projections: 8%

• Investment portfolio data: 5%

Real Scenarios

The HR Summary

An HR manager uses an AI assistant to summarize employee performance reviews before a meeting. Those reviews — containing salaries, disciplinary actions, and personal assessments — now exist on a third-party server with unknown retention policies.

The Code Review

A developer pastes proprietary source code into an AI tool to help debug an issue. That code — potentially containing API keys, business logic, and security vulnerabilities — is now training data.

The Legal Brief

A paralegal uploads a confidential merger document to get a summary. Deal details, financial terms, and negotiation strategies are now outside your data boundary.

When Shadow AI Becomes a Breach Vector: The Vercel Incident

In April 2026, Vercel (the cloud platform behind millions of web deployments) disclosed a significant security incident. The attack didn’t start with a phishing email or a zero-day exploit. It began with an employee using an AI productivity tool.

Here’s what happened:

1. A Vercel employee connected Context.ai, a third-party AI assistant, to their Google Workspace account

2. Context.ai itself was compromised by attackers

3. The attackers leveraged Context.ai’s access to hijack the employee’s Google Workspace

4. From Google Workspace, they pivoted into Vercel’s internal systems

5. Environment variables — including API keys, tokens, and database credentials — were exposed

6. Reports suggest the attackers demanded $2 million in ransom

Vercel’s security team described the attackers as “highly sophisticated based on their operational velocity and detailed understanding of Vercel’s systems.”

The critical insight: the employee did nothing obviously wrong. They used an AI tool to be more productive. They granted it the permissions it asked for. They had no way of knowing Context.ai’s infrastructure would be compromised.

This is the shadow AI risk made concrete. It’s not about employees doing malicious things. It’s about the expanding attack surface that comes from every unauthorized AI integration — each one a potential entry point that security teams can’t see, monitor, or protect.

Why Traditional Security Doesn’t Catch This

Here’s the problem: the employee has legitimate access to all this information. They’re authorized to read HR files, view source code, and handle legal documents.

Traditional Data Loss Prevention (DLP) tools focus on unauthorized access — someone trying to reach files they shouldn’t see. But shadow AI involves authorized users sharing authorized data through unauthorized channels.

The agent has legitimate credentials. The actions look normal. The data moves out through a web browser or API call that’s indistinguishable from regular work activity.

In the Vercel breach, the employee’s OAuth grant to Context.ai looked like any other productivity integration. There was no alert, no flag, no indication that a future supply chain compromise would turn that permission into a breach vector.

The Leadership Problem

This isn’t just a rank-and-file issue. Research has revealed a clear pattern:

• Some C-suite executives prioritize speed over privacy when adopting AI
• Other directors and senior VPs do the same
• Junior employees are more skeptical of this view

When leadership models this behavior, it cascades through the organization.

The Governance Gap

Most companies lack formal AI governance policies. Most employees haven’t received formal AI training. And the majority say they lack clear guidance on what’s acceptable.

The result: employees make individual risk calculations with incomplete information, and security teams can’t secure what they don’t know exists.

The Cost When It Goes Wrong

Organizations with high shadow AI usage face significant financial consequences when breaches occur. The combination of incident response costs, credential rotation, customer notification, regulatory penalties, and reputational damage creates substantial exposure.

Vercel hasn’t disclosed the full financial impact of their April 2026 breach. But for a company that hosts critical infrastructure for thousands of businesses, the costs extend far beyond the immediate incident — customer trust, enterprise sales cycles, and competitive positioning are all affected.

What This Means for Your Organization

The Vercel incident reveals that we don’t just need to worry about employees and colleagues leaking data to AI tools. We should be concerned about those AI tools becoming attack vectors themselves.

Every OAuth grant to an AI service is a trust relationship. Every browser extension that reads your documents is an integration. Every employee “just trying to be more productive” is potentially expanding your attack surface in ways that don’t show up on any security dashboard.

Shadow AI isn’t just a data loss problem. It’s a supply chain security problem. And most organizations have no visibility into either dimension.

In Part III, we’ll look beyond shadow AI to the emerging threats that security researchers are already demonstrating: self-replicating AI worms, persistent compromise mechanisms, and the full “kill chain” of AI agent attacks.

Sources

Shadow AI Usage Statistics:

• UpGuard, “The State of Shadow AI” (November 2025): upguard.com/resources/the-state-of-shadow-ai

• BlackFog Research on Shadow AI (January 2026): blackfog.com/blackfog-research-shadow-ai-threat-grows

Enterprise AI Data Exposure:

• Harmonic Security, “What 22 Million Enterprise AI Prompts Reveal About Shadow AI in 2025”: harmonic.security/resources/…

Vercel Security Incident:

• Vercel Security Bulletin (April 2026): vercel.com/kb/bulletin/vercel-april-2026-security-incident

This is Part II of a three-part series on AI agent security fundamentals.

• Part I: What AI Agents Actually Are (And Why Security Teams Should Care)

• Part II: Shadow AI — The Security Risk Hiding in Plain Sight

• Part III: Emerging Threats — From Prompt Injection to AI Worms

Hundreds of AI papers hit arXiv every day — let alone the breakthroughs coming out of NeurIPS, CVPR, ICLR, MICCAI, ...etc. And that’s before the flood of new repos on GitHub and models on HuggingFace. Nobody can keep up.

Distill AI does the reading for you — it surfaces the papers, code, and models that matter for your specific research, ranked and delivered to you every morning!

👉 Try it now: https://www.distillai.ai

If you are a clinician or researcher, you already know the feeling — hundreds of new papers drop every day across PubMed, medRxiv, ClinicalTrials.gov, and the top journals, and no human has time to keep up.

Distill Medicine fixes that.

It is a personalized daily feed that pulls the latest medical research from across the literature, scores each paper on relevance, journal prestige, recency, and citations, and serves you only what matters to your specialty. Cardiologist focused on heart failure? That’s your feed. Neuroradiologist who never wants to see another mammography paper? Done.

What you get:

• 🎯 26 specialties, 100+ subspecialties — refined to your exact practice

• 📝 Plain-English summaries so you get the idea of any study in 10 seconds

• 📅 Live conference deadline countdowns (RSNA, ASCO, AHA, ASH…)

• 🔖 Bookmarks, notes, and a daily/weekly inbox digest

In an age where staying current is basically a second job, Distill Medicine hands you back your mornings.

👉 Start your free feed at distillmedicine.com

Distill Medicine

SUMMARY

AI agents are fundamentally different from chatbots. They act autonomously, use tools, and persist across sessions. This creates new security risks that traditional approaches don’t address. Understanding the basic threat landscape is the first step toward protecting yourself and your organization.

Want to go deeper? The AI Security Action Pack includes 15 practical guides covering the most common AI agent security risks, plus installable skills your agents can use to protect themselves. Each guide maps to the top 10 most critical AI threats identified by the security organization, OWASP.

Download the AI Security Action Pack

THE CHATBOT VS. AGENT DISTINCTION

Many people just learning about agentic AI believe agents are just smarter chatbots. Here’s the difference:

Chatbots

• Respond to questions
• Stateless (each conversation starts fresh)
• Text in, text out
• You control every interaction

AI Agents

• Pursue goals independently
• Can have persistent memory, spanning days, months and even years of inputs and outputs
• Use tools: web searches, email, file editing, accessing APIs, creating and executing code
• They take actions on your behalf

When you ask ChatGPT “What’s the weather?”, it tells you. When you tell an AI agent “Book me a flight to Denver next Tuesday,” it searches flights, compares prices, enters your payment info, and confirms the booking — potentially across multiple websites, using your credentials.

That’s not a smarter chatbot. That’s delegation of authority.

THE THREE CAPABILITIES THAT CHANGE EVERYTHING

AI agents have three capabilities that chatbots lack. Each one introduces new security considerations:

1. Tool Use
Agents can send emails, read files, execute code, make API calls, and interact with external services. Every tool is a potential attack surface.

2. Memory
Agents remember context across sessions. This enables sophisticated workflows, but it also means malicious instructions can persist. Put harmful instructions into an agent’s memory once, and the attack can survive indefinitely, and influence the agent’s actions over a long period of time.

3. Autonomy
Agents make decisions without checking with you at every step. This is the whole point, but it means an attacker who manipulates the agent’s reasoning can trigger actions you never authorized.

Download the AI Security Action Pack

THE TRUST INVERSION PROBLEM

Here’s the security insight many people miss:

When you use a chatbot, you’re trusting the AI provider. When you deploy an AI agent, you’re trusting everything the agent reads.

Your agent ingests emails, documents, web pages, calendar invites, and code repositories. Any of these can contain hidden instructions that manipulate the agent’s behavior. This is called indirect prompt injection, and it’s the entry point for most AI agent attacks.

The fundamental vulnerability: LLMs process all input as undifferentiated sequences of tokens. There’s no architectural boundary between “trusted instructions from you” and “untrusted content from that email attachment.”

Download the AI Security Action Pack

HIGH-LEVEL THREAT CATEGORIES

For those new to AI agent security, the threat landscape breaks down into three categories:

Data Exposure
Your agent has access to sensitive information. Attackers can trick it into revealing that information through legitimate-seeming requests or by embedding exfiltration instructions in documents the agent processes.

Unauthorized Actions
Your agent can send emails, modify files, and make API calls. Attackers can manipulate the agent into performing actions you never intended, using your credentials and your permissions.

Prompt Manipulation
Attackers craft inputs that override the agent’s intended behavior. This can happen directly (malicious user input) or indirectly (malicious content in files the agent reads).

WHAT THIS MEANS FOR YOU (OR YOUR ORGANIZATION)

If you’re deploying AI agents, or if you or your colleagues are using them for work purposes (without your organization’s knowledge), you need to think about security differently than you would for traditional software.

The attack surface isn’t just the agent itself. It’s every piece of content the agent touches, every tool it can access, and every decision it makes autonomously.

In Part Two, I’ll look at a specific threat that’s already affecting most organizations: shadow AI. This is unauthorized AI tools people are using at work right now, with access to highly sensitive data.

SERIES OVERVIEW

• Part I: AI Agents: The Big Opportunity — and Security Nightmare
• Part II: Shadow AI: The Security Risk Hiding in Plain Sight
• Part III: Emerging AI Security Threats — From Prompt Injection to AI Worms

👉 Go deeper into AI agent security with our free Action Pack.

If you work in AI, you know this feeling.

You open arXiv on a Monday morning. 500+ new papers. You skim titles for 30 minutes. You bookmark a few. You miss the important ones. Then you check Twitter, Google Scholar, three newsletters, and a Slack channel.

Your morning is gone. And you’re still not caught up.

I’ve been living this problem for years — as a research fellow working on medical imaging and deep learning, keeping up with the literature became a second job.

So, we’ve go you something to fix it.

Introducing Distill AI

distillai.ai is a personalized daily AI research feed that pulls from arXiv, Nature, NeurIPS, CVPR, ICML, ICLR, GitHub, and HuggingFace — then scores and ranks everything to your specific research interests.

Instead of scrolling through hundreds of papers, you get the ones that actually matter to you. Every morning. In your feed and your inbox.

Distill AI

How it works

Pick your topics. Choose from 54 AI specializations — computer vision, LLMs, reinforcement learning, medical imaging, robotics, and more.

We do the rest. Our scoring algorithm evaluates every new paper based on code availability, lab reputation, recency, impact signals, and relevance to your topics. The highest-scored papers land in your feed.

Everything in one place. Not just papers — trending GitHub repositories, HuggingFace models, and conference deadline countdowns (NeurIPS is 27 days away). All personalized to what you care about.

Distill AI

What’s included

🔬 Personalized feed — papers matched to your interests, refreshed daily
📊 Relevance scoring — our algorithm surfaces the most impactful work
📝 2-sentence summaries — understand any paper instantly
💻 GitHub & HuggingFace — trending repos and models in your topics
🔖 Bookmarks & notes — save papers with your personal research notes
📬 Inbox delivery — daily or weekly digest sent to your email
⏰ Deadline tracker — live countdowns for NeurIPS, EMNLP, AAAI, ICLR, and more

Distill AI

Free to start

You can try it right now — no credit card needed. Free gives you 3 papers/day and 1 topic. Pro unlocks up to 50 papers/day, all 54 topics, summaries, bookmarks, and email delivery.

The first 500 Pro members lock in a founding price of $4.99/month (50% off the regular price — forever).

👉 Try Distill AI for free → distillai.ai

Legal tech is evolving from back office support to frontline enforcement.

While most AI headlines focus on generative tools for coding, writing, or design, a quiet revolution is happening in a very different part of the tech world. Darrow, a legal tech company, has built technology that helps lawyers find legal violations that no one else sees.

This new category of legal tech is called legal intelligence, and it’s the structured process of scanning public data to detect legal violations at scale. Instead of waiting for harm to be reported, carefully built legal intelligence infrastructure finds patterns of misconduct hiding in publicly available sources and builds evidence to turn them into actionable cases.

This is exactly what Darrow is doing. Combining large language models, machine learning, web intelligence, and legal expertise, this tech company is powering a new kind of legal discovery.

Most Data Gets Ignored. Darrow Turns It Into Evidence.

Public data is published every second across the internet: user complaints, app store reviews, environmental filings, financial disclosures, and regulatory records. These documents often sit untouched, even when they contain signs of unlawful behavior.

Darrow’s technology scans data from thousands of sources, across multiple domains, and AI models summarize and organize unstructured text, then cluster similar issues together. The team’s attorneys and legal intelligence analysts then evaluate whether the content might point to a violation of law.

This process blends automation with human judgment. It doesn’t replace lawyers, but it gives them a clearer view of where harm is happening and how to respond.

For example, Chicago-based class action attorney, Kristina Carroll, partnered with Darrow after the company flagged privacy concerns involving the dating apps Bumble and Badoo. Carroll filed a class action against parent company Bumble Inc., which settled last year for $40 million.

What’s fascinating is the fact that the case didn’t come from a whistleblower or a high-profile news investigation, but from public data interpreted by attorneys and a legal intelligence system trained to look for the kinds of issues most people miss.

A Targeted System for Different Types of Harm

Not every legal violation looks the same. A case about data privacy involves different signals than one about environmental harm or deceptive pricing. Darrow’s team of data scientists, engineers, attorneys, and legal intelligence analysts build specialized detection tools for each area of law, combining subject-matter expertise with training data specific to the problem.

The company builds tools with different intent, based on the type of violation it’s looking for. In an antitrust investigation, the company might focus on price changes and customer complaints. In a privacy case, it might flag suspicious onboarding language or metadata collection.

Legal Intelligence Makes the Invisible Visible

Evidence for legal violations is often scattered across unrelated sources, buried in technical documents, or locked behind vague legalese.

Legal intelligence brings that evidence to the surface. It helps lawyers spot systemic problems, evaluate them faster, and take action before harm spreads further.

Darrow is building the infrastructure needed to uncover violations at scale and connect them to real-world enforcement, creating a whole new approach to scaling legal accountability.

To learn more about Darrow’s technology, check out this article on Darrow’s blog: Using Legal Intelligence to Turn Data Into Justice

Hello, dear readers!

This edition covers major AI developments that are making headlines—a dramatic clash between the U.S. government and a leading AI company, a surprising revelation about how top engineers now work, and travel booking getting a conversational upgrade. Let’s dive in!

🏛️ Trump Orders Federal Government to Cut Ties With AI Company Anthropic

President Trump ordered all federal agencies to immediately stop using Anthropic’s technology after a heated standoff with the Pentagon. Anthropic refused to allow unrestricted military use of its AI, seeking guarantees it wouldn’t be used for autonomous weapons or mass domestic surveillance. Defense Secretary Pete Hegseth designated Anthropic a “supply chain risk”—a label typically reserved for foreign adversaries—barring military contractors from doing business with the company. Anthropic has vowed to challenge the designation in court, calling it “legally unsound.”

💻 Spotify Says Its Top Developers Haven’t Written Code Since December—AI Is Doing It For Them

Spotify co-CEO Gustav Söderström revealed that the company’s best developers “have not written a single line of code since December.” Instead, engineers use an internal AI system called “Honk” built on Claude Code to generate and deploy code—even from their phones via Slack during their morning commute. The system helped Spotify ship over 50 new features in 2025. Söderström called this “just the beginning” of AI-driven development at the company.

🏠 Airbnb Is Testing AI-Powered Search to Help Users With Bookings

Airbnb is testing a new AI-powered search feature that lets users describe what they’re looking for in natural language instead of using traditional filters. Users can type prompts like “a pet-friendly cabin with a hot tub near the mountains” and ask follow-up questions about listings. CEO Brian Chesky said the company is building an “AI-native experience where the app doesn’t just search for you—it knows you.” The feature is currently live for a small percentage of users.

As the demand for high-speed connectivity continues to grow, mobile operators are struggling to keep up. Their current approach—expanding Radio Access Networks (RANs)—is not sustainable. RANs account for 80% of all mobile network energy consumption, with energy costs outpacing sales growth by over 50%.

Recognizing this growing challenge, Net AI was founded to develop AI-driven solutions that address network efficiency and sustainability. The mission is to bridge the gap between telecommunications and machine learning, enabling operators to optimize the performance of their networks while reducing energy consumption.

Industry Status Quo

Most actors in the telecommunications industry believe that AI systems are not accurate enough in their predictions of traffic demands and would consume too much energy themselves. This stems from a disconnect between the telecom and machine learning (ML) communities.

Traditionally, telecom networks have relied on rigid, rule-based optimization techniques designed for deterministic systems, while ML involves building statistical models based on extensive data used for training. Because of this fundamental difference, many telecom professionals are skeptical about allowing AI to make decisions that could impact network reliability. On the other hand, ML experts often underestimate the strict performance requirements and real-time constraints specific to telecom networks. This disconnect has prevented the industry from fully realizing AI’s potential in optimizing network operations. Bridging this gap requires specialized AI models tailored to telecom environments, ensuring both accuracy and efficiency without compromising service quality.

Meet Net AI

Net AI has been developing domain-specific ML/AI models that can help telecom operators reduce energy consumption and improve network efficiency. Our research has shown that significant energy savings—up to 60%—can be achieved by selectively deactivating parts of the network infrastructure during low-demand periods. However, this must be done carefully to ensure that service quality is not compromised.

Its approach focuses on:

📡 Accurate Traffic Forecasting: Providing real-time insights into network performance and resource usage to enable informed decision-making.

⚡ Efficiency Improvements: Helping operators optimize network resources to manage energy costs while maintaining service quality.

🔎 Proactive Network Management: Identifying unusual patterns in traffic to predict and prevent network incidents.

Our team’s unique expertise at the intersection of AI and telecoms has enabled us to build perhaps the most accurateAI-driven forecasting engine on the market (patent pending). What makes it different?

🚀 Scalability: Our AI is 100 times more scalable than competitors and inherently low in power consumption. We have achieved this via domain-specific neural architecture design.

🎯 Performance: We achieve over 95% accuracy in traffic forecasting. Our data handling pipeline employs pre-processing that improves the sensitivity and generalization ability of trained models.

🔒 Reliability: Our proprietary training methodology ensures that AI models avoid underestimation errors, crucial for maintaining service quality.

By challenging the widely held perception that service disruption risks and computational cost outweigh the benefits of AI-driven forecasting, we’ve created a solution that gives operators confidence to automatically and intelligently switch off targeted network infrastructure while maintaining service quality.

With additional integration work, similar strategies could be applied in data centers, addressing energy efficiency challenges on a broader scale.

Shaping the Future of Telecoms 🌍📶

The telecommunications industry is at a critical juncture where intelligent automation and energy efficiency are becoming key priorities. As AI continues to evolve, its role in infrastructure management will expand, providing new opportunities to enhance service quality while reducing operational costs.

Net AI is playing an active role in this transformation. By combining expertise in AI and telecom, they aim to contribute to a more efficient, intelligent, and sustainable future for network operations. Their work within the ETHER project further supports this goal as it focuses on improving network coverage and energy efficiency through integrated terrestrial and non-terrestrial networks.

Net AI has already demonstrated that AI-driven solutions can play a meaningful role in optimizing telecom infrastructure. As the industry moves forward, we continue to explore new ways to refine and expand these capabilities.

HackerRank, a company primarily known for its tech hiring solutions, recently introduced a benchmark designed to challenge AI models and test how well LLMs solve complex software tasks. The initial press release explained, “The ASTRA Benchmark consists of multi-file, project-based problems designed to mimic real-world coding tasks. The intent of the HackerRank ASTRA Benchmark is to determine the correctness and consistency of an AI model’s coding ability in relation to practical applications.”

Seeming to acknowledge the deviation from the company’s other offerings, which focus on screening, interviewing, engaging and skilling developer talent, HackerRank co-founder and CEO Vivek Ravisankar said, “As software development becomes more human + AI, it’s important that we have a very good understanding of the combined abilities. Our experience pioneering the market in assessing software development skills makes us uniquely qualified to assess the abilities of AI models acting as agents for software developers.”

Whether true or not, HackerRank ASTRA enters the market at an interesting moment, with a battle royale raging between AI2, Anthropic, DeepSeek, Google, OpenAI, xAI and others. While the initial release of HackerRank ASTRA does not address all aspects of these models, it does acknowledge the limitations of other benchmarks and seeks to offer a more comprehensive framework, according to the paper published by the HackerRank team.

As the paper explains, “HackerRank-ASTRA is a benchmark built from HackerRank’s proprietary library of multi-file, project-based software development problems. These problems were originally designed to assess the software development skills of human developers across a wide range of skill domains in realistic, project-like settings. We observed that even advanced large language models (LLMs) face significant challenges when solving these problems, which motivated the creation of this benchmark.”

As such, the current version of the open-source benchmark includes:

• Diverse skill domains: Sixty-five product-based coding questions, categorized into 10 primary coding skill domains and 34 subcategories.

• Multi-file project questions: An average of 12 source code and configuration files per question as model inputs. This results in an average of 61 lines of solution code per question.

• Model correctness and consistency evaluation: HackerRank ASTRA prioritizes metrics such as average scores, average pass@1 and median standard deviation.

• Wide test case coverage: HackerRank ASTRA’s dataset contains an average of 6.7 test cases per question to rigorously evaluate the correctness of implementation.

• Leaderboard: Full report and analysis results are available on the HackerRank website.

There are some limitations to the initial benchmark, which HackerRank acknowledged in its research paper. Notably, the benchmark focuses on front-end projects, which means it under-represents back-end skills and other domains. Still, HackerRank is already on top of that issue, with plans for future iterations that will include a broader range of technologies. On top of that, newer AI models are still being added to the leaderboard (about as quickly as they enter the market). Usefulness and application over time are ultimately up to the AI community to determine, but for now, HackerRank points to what the benchmark actually is proving, and that’s how today’s frontier models perform on real software development challenges.

This edition brings you critical developments reshaping the AI landscape—from America’s infrastructure challenges to job disruption warnings and Microsoft’s bold move against Nvidia. Let’s dive in!

🏗️ NVIDIA’s CEO Has a Blunt Warning About America’s AI Infrastructure Problem

Jensen Huang warned that the US faces a growing infrastructure gap with China. While American data centers take three years to build, China “can build a hospital in a weekend.” He also noted China has twice the energy capacity despite a smaller economy. Though NVIDIA remains “generations ahead” in chip technology, Huang cautioned that infrastructure bottlenecks could undermine America’s AI leadership.

💼 The Man Who Helped Build AI Just Issued His Bleakest Warning Yet About Your Job

Anthropic CEO Dario Amodei predicts AI could eliminate 50% of entry-level white-collar jobs within 1-5 years, pushing unemployment to 10-20%. Tech, finance, law, and consulting sectors face the biggest disruption. His message: “You can’t step in front of the train and stop it. The only move that’s going to work is steering the train.”

🔧 Microsoft’s Secret Weapon: The Chip That Could Dethrone Nvidia in the AI Wars

Microsoft unveiled the Maia 200, its second-generation AI chip built to challenge Nvidia. The chip is 30% cheaper, uses far less power (750W vs Nvidia’s 1,200W+), and outperforms Amazon and Google’s custom chips in key benchmarks. Microsoft’s long-term goal: “mainly Microsoft silicon in the data center.”

Menu-Order-AI was created to solve a very real, very current problem: people love going out to eat, but they’re ordering differently than they used to. With nearly 9% of the U.S. population now using GLP-1 medications and millions more eating with intention, the traditional restaurant menu no longer supports how diners actually make decisions.

Diners aren’t asking restaurants to change their menus. They still want to enjoy the experience, the atmosphere, and the social moment. What they need is confidence — confidence that what they’re ordering aligns with how they eat right now. That moment of uncertainty happens at the table, not before they leave the house and not after the meal is over.

Unlike calorie trackers or meal planners that work before or after eating, Menu-Order-AI operates in real time. From one restaurant choice or scan of a menu, the app instantly analyzes the menu and highlights options that align with high-protein, GLP-1 friendly, low-carb, low WW point, and plant-based preferences. No guesswork. No jumping between apps. No awkward ordering conversations.

This matters because GLP-1 users aren’t eating less just to eat less — they’re prioritizing quality, protein, and how food makes them feel. Restaurants are already seeing this shift reflected in smaller portions ordered, fewer shared appetizers, and lower alcohol consumption. Industry data shows restaurants losing an estimated 7% of revenue due to these changes, not because people stopped dining out, but because they’re unsure what to order.

By guiding diners toward meals that work for their lifestyle, the app restores confidence at the table. When diners feel confident, they enjoy the experience more — and they’re more likely to return. The solution doesn’t disrupt restaurants, require menu changes, or add friction. It simply supports the diner at the exact moment of decision.

This is why Menu-Order-AI is best described as a real-time dining companion, not a diet app. It’s designed for real life: business dinners, date nights, travel, group meals, and everyday dining — whether eating in or ordering delivery.

As GLP-1 adoption continues to rise and intentional eating becomes the norm, the question is no longer if dining behavior has changed, but how restaurants and diners adapt together. Menu-Order-AI bridges that gap by empowering diners to order with confidence — every time, anywhere.

The app is available now on the Apple App Store and Google Play Store.

Menu-Order-AI

⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐
Would you like to be featured in our newsletter🔥 and get noticed, QUICKLY 🚀 (70,000+ subscribers)? Simply reply to this email or send an email to editor@aibuzz.news, and we can take it from there.
⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐

Hey AI enthusiasts! 👋

This week’s edition is packed with seismic shifts in the AI landscape. We’re talking billion-dollar chip deals, unexpected comebacks, and trade policies that could reshape the entire industry. From OpenAI’s game-changing hardware partnership to Google proving the doubters wrong, plus a tariff twist that has Nvidia celebrating — buckle up, because the AI world just got a whole lot more interesting. Let’s dive in! ⚡

The Chip That Could Change Everything: Inside OpenAI’s Cerebras Deal

OpenAI just locked in a massive $10 billion partnership with Cerebras Systems to supercharge its AI infrastructure. The deal will deploy 750 megawatts of computing power through 2028, making it the largest high-speed AI inference deployment ever. Cerebras’ unique wafer-scale chips promise responses up to 15x faster than traditional GPUs — a critical upgrade as OpenAI shifts from chatbots to autonomous AI agents where every millisecond counts.

The Empire Strikes Back: How Google Went from Underdog to AI Champion

Remember when Google was in “code red” panic mode after ChatGPT launched? Those days are over. The release of Gemini 3 sent shockwaves through the industry, with Salesforce’s CEO declaring he’s never going back to ChatGPT. Google’s secret weapon? An unmatched ecosystem spanning Android, Search, Gmail, and YouTube — processing 5x more AI tokens than OpenAI monthly. The sleeping giant is fully awake.

Pay to Play: Trump Opens the Door to China — For a Price

President Trump just flipped the script on AI chip exports with a 25% tariff on sales to China — and surprisingly, Nvidia is cheering. The new “revenue-capture” model reopens the Chinese market (previously banned) while funneling billions to the U.S. government. The catch? Chips for domestic data centers and research are exempt, creating a powerful incentive to build AI infrastructure on American soil. Intel and TSMC are already positioning for the reshoring wave.

Until next time, stay curious! 🤖✨

Apply to PharmStars Spring Cohort

Are you a digital health startup eager to sell to pharma? Don’t miss the opportunity to participate in PharmStars’ Spring accelerator focused on “Digital Innovations in Operations & Outcomes.”

PharmStars provides startups with: 

🔹 In-depth pharma education through structured PharmaU^TM curriculum

🔹 Personalized mentorship from former pharma executives

🔹 Access to pharma innovation leaders and decision-makers 

🔹 Opportunities to pitch to potential pharma clients

In addition, startup graduates receive an investment from the PharmStars Ventures fund.

What are you waiting for?

APPLY TODAY

Learn more and apply by January 25

Like most industries, AI and tech have had a profound effect on the legal industry. For years, most legal tech tools focused on making existing processes faster and more organized. However, what’s emerging are more advanced tools that point to a broader shift in how attorneys do their work. Instead of focusing purely on efficiency, legal tech is enabling a process level change where attorneys can rely on technology to support entire legal workflows.

Here are three trends we’re seeing in the legal tech arena as we go into 2026.

Darrow AI

#1 Litigation analytics becomes a core capability

Litigation analytics will function as foundational infrastructure rather than a specialized feature. Earlier generations of analytics concentrated on historical outcomes, but in 2026, analytics will operate continuously and earlier in the lifecycle, shaping how matters are identified and evaluated.

Advances in large scale analysis of litigation activity and public data will enable earlier detection of potential claims and exposures. As new information enters these systems, analytical models will update assessments of case strength and litigation risk throughout the life of a matter. These outputs are likely to influence portfolio construction and capital allocation rather than serve as static reference material.

The same analytical foundation is expected to support adjacent domains such as insurance underwriting and litigation finance. Filed cases will increasingly represent one expression of a broader intelligence layer that informs valuation and risk.

Darrow AI

#2 The industry moves toward agent-based models

Legal AI in 2026 will rely more heavily on agent-based systems capable of executing multistage work with limited human input. Instead of isolated prompt interactions, platforms will deploy coordinated agents that carry out sequences of analysis as conditions evolve.

These agents are likely to operate in parallel across distinct responsibilities. One may handle docket ingestion and normalization. Another may focus on jurisdictional dynamics or precedent analysis. Additional agents may evaluate exposure and downstream implications. Working together, they will support complex legal workflows and enable higher quality decisions at scale.

Agent-based systems align closely with the iterative and context dependent nature of legal work. They’re going to allow human attorneys to focus more on strategy and judgment rather than manual synthesis.

Darrow AI

#3 Embedded intelligence replaces standalone legal tech tools

Legal intelligence is expected to surface contextually within existing systems rather than through standalone dashboards. Insight will appear at the point of decision, embedded directly into operational workflows.

Legal and regulatory risk assessment is likely to become part of routine operations, and legal violations, such as data breaches, will be detectable using public data before products or features reach market.

As intelligence is designed to function across multiple environments, operational friction will decrease while access expands. Firms will rely on fewer platforms and expect deeper analytical value from each. Intelligence will operate as an execution layer integrated into daily work, supporting action rather than requiring deliberate consultation.

Darrow is the leading legal intelligence company identifying legal risk at scale. To learn more about how Darrow uncovers legal violations, check out this article on Darrow’s blog: Using Legal Intelligence to Turn Data Into Justice.For more information about Darrow, visit www.darrow.ai.

⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐
Would you like to be featured in our newsletter🔥 and get noticed, QUICKLY 🚀 (71,000+ subscribers)? Simply reply to this email or send an email to editor@aibuzz.news, and we can take it from there.
⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐

HackerRank, a company primarily known for its tech hiring solutions, recently introduced a benchmark designed to challenge AI models and test how well LLMs solve complex software tasks. The initial press release explained, “The ASTRA Benchmark consists of multi-file, project-based problems designed to mimic real-world coding tasks. The intent of the HackerRank ASTRA Benchmark is to determine the correctness and consistency of an AI model’s coding ability in relation to practical applications.”

Seeming to acknowledge the deviation from the company’s other offerings, which focus on screening, interviewing, engaging and skilling developer talent, HackerRank co-founder and CEO Vivek Ravisankar said, “As software development becomes more human + AI, it’s important that we have a very good understanding of the combined abilities. Our experience pioneering the market in assessing software development skills makes us uniquely qualified to assess the abilities of AI models acting as agents for software developers.”

Whether true or not, HackerRank ASTRA enters the market at an interesting moment, with a battle royale raging between AI2, Anthropic, DeepSeek, Google, OpenAI, xAI and others. While the initial release of HackerRank ASTRA does not address all aspects of these models, it does acknowledge the limitations of other benchmarks and seeks to offer a more comprehensive framework, according to the paper published by the HackerRank team.

As the paper explains, “HackerRank-ASTRA is a benchmark built from HackerRank’s proprietary library of multi-file, project-based software development problems. These problems were originally designed to assess the software development skills of human developers across a wide range of skill domains in realistic, project-like settings. We observed that even advanced large language models (LLMs) face significant challenges when solving these problems, which motivated the creation of this benchmark.”

As such, the current version of the open-source benchmark includes:

• Diverse skill domains: Sixty-five product-based coding questions, categorized into 10 primary coding skill domains and 34 subcategories.

• Multi-file project questions: An average of 12 source code and configuration files per question as model inputs. This results in an average of 61 lines of solution code per question.

• Model correctness and consistency evaluation: HackerRank ASTRA prioritizes metrics such as average scores, average pass@1 and median standard deviation.

• Wide test case coverage: HackerRank ASTRA’s dataset contains an average of 6.7 test cases per question to rigorously evaluate the correctness of implementation.

• Leaderboard: Full report and analysis results are available on the HackerRank website.

There are some limitations to the initial benchmark, which HackerRank acknowledged in its research paper. Notably, the benchmark focuses on front-end projects, which means it under-represents back-end skills and other domains. Still, HackerRank is already on top of that issue, with plans for future iterations that will include a broader range of technologies. On top of that, newer AI models are still being added to the leaderboard (about as quickly as they enter the market). Usefulness and application over time are ultimately up to the AI community to determine, but for now, HackerRank points to what the benchmark actually is proving, and that’s how today’s frontier models perform on real software development challenges.

Over the summer, Darrow, the legal intelligence company built to surface hidden legal risk and turn it into action, detected an anomaly in the code of a fertility tracking app.

On the surface, nothing seemed unusual. The app’s latest update notes mentioned only “general performance improvements” and “minor bug fixes.” But Darrow’s legal intelligence infrastructure flagged something unexpected: the app started transmitting unique device identifiers to an advertising partner that was not listed anywhere in its privacy policy.

If true, this would mean that users’ extremely sensitive reproductive data is unknowingly being tracked and profiled to a third party without consent.

Darrow AI

At first, the signal did not appear concerning. It was the kind of background telemetry many apps use for crash analytics or ad attribution. Still, Darrow’s system marked it as suspicious. The Legal Intelligence team compared the app’s new code signature to its earlier version, then ran the update through a series of checks: changes in requested permissions, modifications to background processes, and differences in outbound network traffic. This revealed a significant change and potential privacy violation: the app had quietly integrated a new module from an advertising software development kit (SDK) that, according to publicly available documentation, could link identifiers with precise geolocation data and trace users’ routines from morning check-ins to evening updates.

Darrow AI

To the average user, this would be invisible. From a legal perspective, however, it raised serious concerns. Analysts reconstructed the data flow, cross-referencing the app’s behavior with recent rulings involving mobile tracking technologies. Combining persistent identifiers and location data and transmitting them to a third party without explicit consent can violate several privacy statutes, including the California Invasion of Privacy Act (CIPA), a state statute that prohibits the unauthorized interception or sharing of communications and data, and often serves as the basis for class action lawsuits.

Darrow automatically flagged the incident as a potential privacy violation and sent it to the company’s privacy intelligence team for legal review. The privacy team began investigating and tracked the timeline of the code changes and confirmed that the new permissions requested access to background location data even when the app was not running. Legal analysts reviewed the information alongside recent court rulings involving unauthorized tracking in mobile applications. The team’s review showed a strong resemblance to earlier cases that resulted in major settlements, particularly those where users had no way to opt out.

Over the following days, Darrow’s legal intelligence analysts mapped the full scope of the issue and analyzed public download data and app store activity to estimate that millions of users had installed the affected version since the update. Each of them likely had their movements and device IDs logged and shared with an undisclosed third party.

Because the data was highly sensitive and the exposure widespread, the incident constituted a serious privacy breach. Darrow’s legal experts identified this as an alleged CIPA violation and translated the findings into a full assessment, including projected damages and settlement scenarios. They assembled a detailed case memorandum with the evidence, legal theories, and impact analysis, and brought it to a leading plaintiff firm. The firm determined the case was viable and filed a putative class action to seek relief for the exposed users.

Through the combination of machine learning, data analysis, and legal expertise, Darrow’s legal intelligence analysts turned a few lines of suspicious code into a potential nationwide enforcement action that could secure relief for millions of people whose most intimate health choices were tracked and sold without their consent.

If successful, the case will send a clear message that exploiting reproductive data for ad targeting carries real financial and legal consequences. This is the power of legal intelligence: turning invisible technical changes into accountability, and shifting power back to the people whose lives and data are on the line.

To learn more about how Darrow uncovers legal violations, check out this article on Darrow’s blog: Using Legal Intelligence to Turn Data Into Justice. For more information about Darrow, visit www.darrow.ai.

Darrow AI