Skip to main content

The Nexairi Dispatch  ·  Friday, April 24, 2026  ·   Issue #6

271 bugs found: AI is better at hacking than defense

Anthropic's Mythos found 271 Firefox zero-days — 12x more than Claude Opus. Mozilla is patching fast.

By Jim Smart

Good morning, friends. Anthropic's private security model just found 271 zero-day vulnerabilities in Firefox — twelve times more than what standard Claude found in the same test, and a number that should make every security team pause. OpenAI made ChatGPT free for every verified U.S. clinician this week, removing the last cost barrier to AI in the clinic. And SpaceX negotiated a $60 billion option to buy an AI coding startup, with a $10 billion walk-away clause if they change their minds. The platform wars are getting expensive.


🔒 AI SECURITY — Anthropic's AI found 271 Firefox zero-days

What happened: Anthropic's restricted Mythos model found 271 zero-day security vulnerabilities in Firefox — 12 times more than Claude Opus 4.6 found in the same test. The discovery came from coordinated security research, and Mozilla is now actively patching the vulnerabilities across affected browser versions.

Why it matters: Purpose-built security AI can outperform general models by a factor of 12 — and that gap is the whole story. The same capability that lets defenders find 271 holes in one run will eventually be in the hands of people who want to exploit those holes. The dual-use problem isn't theoretical anymore.

What to watch: Watch for specialized red-team AI models from other frontier labs, and how policymakers respond when these tools prove too effective to keep quiet.

Read the full analysis →


🏭 TECHNOLOGY — Tesla is building its own AI chip factory in Austin

What happened: Tesla and SpaceX are jointly building TERAFAB, an AI chip fabrication facility in Austin, Texas, with Intel as the manufacturing partner. Tesla's $25 billion 2026 capital expenditure — roughly three times what it spent last year — funds the project. The goal is to produce chips for Optimus robots and Full Self-Driving inference in-house, replacing Nvidia purchases.

Why it matters: Apple's shift from Intel chips to its own M1 in 2020 changed the economics of hardware permanently. Tesla is betting the same logic applies to AI inference at scale. If TERAFAB works, Tesla controls its chip supply chain through 2030 and beyond. If it doesn't, $25 billion sits in Austin.

What to watch: This is Part 1 of a 7-part series tracking Elon's interplanetary stack. Watch whether Intel holds up as the manufacturing partner — and whether TSMC gets cut out entirely.

Read the full analysis →


🏥 HEALTHCARE — ChatGPT is now free for every verified U.S. doctor

What happened: OpenAI opened free ChatGPT access to verified U.S. clinicians — doctors, nurses, and pharmacists — starting April 22. The program removes the $20/month cost barrier and covers drafting clinical notes, summarizing research, and explaining diagnoses in plain language. The move follows Google's competing push to embed AI into hospital workflows.

Why it matters: Healthcare is one of the last major professional verticals where AI adoption stalled partly on cost. Removing the subscription fee for hundreds of thousands of clinicians is a distribution move, not a product launch — and it's how OpenAI wins the clinical workflow race before Google does.

What to watch: Many hospitals maintain blanket bans on patient data entering commercial AI tools. Free access doesn't change those policies, and how clinicians navigate that friction will define actual adoption rates.

Read the full analysis →


💻 TECHNOLOGY — $60B for Cursor — or $10B just to walk away

What happened: SpaceX negotiated a $60 billion acquisition option for Cursor, the AI coding assistant popular with developers and engineers. The deal includes a $10 billion exit clause — SpaceX pays that amount if it decides not to complete the purchase. The Wall Street Journal first reported the terms, confirmed by multiple parties.

Why it matters: A $10 billion walk-away fee is not a sign of uncertainty. It's a signal that SpaceX considers letting a competitor own Cursor to be worth $10 billion to prevent. OpenAI, Anthropic, and Google are all competing for control of the tools developers use daily — whoever owns that layer owns the development stack for the next decade.

What to watch: Whether Cursor's research team stays post-acquisition matters as much as the price. The best coding AI tools depend on research velocity — and that lives in the people, not the product.

Read the full analysis →


Outside Nexairi

GPT-5.5 is out — and it costs twice as much to run — TechCrunch

OpenAI released GPT-5.5 on April 23, just six weeks after GPT-5.4. The model is stronger at multi-step tasks — planning, tool use, code, and scientific workflows — with fewer hallucinations. It rolls out to Plus, Pro, Business, and Enterprise users, and API pricing is 2x higher than GPT-5.4. OpenAI president Greg Brockman called it a new class of intelligence for real work.

Anthropic explained what broke Claude Code for six weeks — Anthropic Engineering

Anthropic published a detailed postmortem on April 23 explaining three separate changes that degraded Claude Code performance from March through April: a reasoning effort default quietly lowered to save latency, a caching bug that erased Claude's working memory mid-session, and a verbosity instruction that cut coding quality by 3%. All three are fixed as of April 20, with usage limits reset for all subscribers.

Meta is logging employee keystrokes and clicks to train AI — MIT Technology Review

Meta is rolling out monitoring software on employee devices to capture clicks and keystrokes, using the behavioral data to improve AI model development. Workers have raised concerns about surveillance scope, and the move highlights how AI labs are turning to internal behavior data to fill gaps that public datasets can't.

Claude Desktop installs a browser bridge it doesn't tell you about — Let's Data Science

Researchers found that Anthropic's Claude Desktop app installs a pre-authorized browser extension bridge without prominently disclosing it during setup, allowing Claude to interact with open browser tabs. The discovery raises questions about what AI desktop apps are actually doing with permission — and what users are consenting to.


Tool Worth Knowing: IFTTT MCP

IFTTT launched a Model Context Protocol connector that links Claude directly to over 1,000 apps — including Gmail, Google Sheets, Slack, and smart home devices. If you're already using Claude for daily work, this is the automation layer that connects your AI to everything else you already use.


Deeper Read

ChatGPT Workspace Agents: What They Do and Risks to Know — Nexairi

Before you let autonomous agents loose on your team's tools, here's the honest breakdown of what they can handle — and where they quietly break down.

AI Built the Code. Now Nobody Can Fix It. — Nexairi

Teams shipping code with AI are quietly accumulating technical debt with no owner — here's how the high-performing teams handle it before it compounds.


Quick Hits

Share this issue

Get every issue in your inbox

3 issues a week — Mondays, Wednesdays, Fridays. Free.

Subscribe free →