In partnership with
A blog post that nobody was supposed to see. In late March, Anthropic accidentally published nearly 3,000 internal documents to the public web. Among them was a draft announcement for an unreleased model called Claude Mythos, described internally as "by far the most powerful AI model we've ever developed." Security researchers at the University of Cambridge and LayerX Security found the exposed files before Anthropic took them down. CrowdStrike dropped 7% that day. Palo Alto fell 6%. An accidental blog post moved markets.
On April 7th, Anthropic made it official. Claude Mythos Preview launched as the centrepiece of Project Glasswing — a restricted cybersecurity initiative giving around 40 organisations controlled access. Microsoft, Apple, Google, Amazon, CrowdStrike, Cisco, JPMorgan Chase and the Linux Foundation were all at the table. Anthropic committed $100 million in usage credits. This was not a normal product launch.
What went down
An accidental blog post exposed Mythos weeks before launch, crashing cybersecurity stocks.
Anthropic describes Mythos as the most powerful — and potentially most dangerous — AI model ever built.
The model can find and exploit software vulnerabilities faster than any human security researcher.
US Treasury summoned bank CEOs to discuss AI-driven cyber risks
During safety testing, it broke out of the secure environment it had been placed in.
Partner message
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
Why does this matter?
Anthropic's own system card flagged a "potentially dangerous capability" to bypass safeguards — a model that can break out of its containment environment is categorically different from one that simply writes good code.
This lands against a backdrop of documented AI-assisted attacks: in September 2025, a Chinese state group used Claude Code to autonomously infiltrate ~30 targets; months earlier, a low-skilled criminal used Claude to build and sell ransomware.
Our take
The counter-argument is fair: "too powerful to release" has become a recurring marketing beat. Restricted access to enterprise clients is a premium sales strategy. The accidental leak, whether truly accidental or not, generated extraordinary publicity. And the benchmark picture at the frontier remains genuinely competitive — GPT-5.4 leads on some coding tasks, Gemini wins on context and cost efficiency.
But the press coverage tells you something. NBC News, CNBC, SecurityWeek, and The Hacker News all treated Mythos as a security story with real-world consequences — not an AI hype cycle. The US Cybersecurity and Infrastructure Security Agency Cybersecurity and Infrastructure Security Agency and the Center for AI Standards and Innovation are in active discussions with Anthropic about its capabilities. JPMorgan Chase doesn't join defensive coalitions over theoretical risks. Whatever you make of the framing, the pattern of the past twelve months — benchmark dominance, documented AI-driven attacks, a model leak that moved markets, and now a formal restricted deployment — has earned more scrutiny than most model announcements.
Another big thing…OpenAI's turbulent week
OpenAI shelved Stargate UK, a centrepiece of the £31bn UK-US AI investment deal announced last September, citing high energy costs and regulatory uncertainty. The decision is a blow to the UK government's AI-led growth strategy.
Separately, OpenAI responded to a supply-chain compromise involving the Axios developer tool by rotating macOS code signing certificates and confirming no user data was affected.
And CEO Sam Altman published a blog post addressing both a molotov cocktail attack on his San Francisco home and allegations raised in a New Yorker profile questioning his leadership. A week in which the company's infrastructure ambitions, security posture and chief executive all came under pressure simultaneously.
