Claude Mythos Preview

Claude Mythos Preview is Anthropic's most capable cybersecurity AI model, released on 8 April 2026 exclusively to vetted partner organisations through Project Glasswing. The model scored 83.1% on the CyberGym vulnerability reproduction benchmark against 66.6% for Anthropic's previous best, and autonomously identified thousands of zero-day vulnerabilities including a 27-year-old flaw in OpenBSD — performance that prompted Anthropic to withhold it from public release. Alongside the Glasswing expansion on 7 April, Anthropic published a 244-page Alignment Risk Update and abandoned ASL (AI Safety Level) capability benchmarks in favour of autonomy-focused threat models.

Tom's Hardware published a critical review arguing the "thousands of zero-days" claim rested on only 198 manual reviews, many in outdated software. The challenge was partially answered by the UK's AI Safety Institute (AISI), which published an independent evaluation on 15 April 2026 confirming a 32-step attack chain equivalent to 20 human work-hours — the first independent verification of Mythos capability. AISI found no single-task superiority, vindicating Tom's Hardware's critique of the headline claim, while the chaining result demonstrated a different and arguably more operationally significant capability.

The US government's reaction was immediate. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell summoned the CEOs of five major banks to an emergency meeting specifically to discuss Mythos — the first time federal financial regulators have convened Wall Street leadership over a single AI model. The AISI evaluation published a week later provides Treasury and the Fed with independent technical grounding they lacked on 8 April, when they were acting on Anthropic's own claims alone.