On April 8, Anthropic released Claude Mythos, an internal frontier model it claims scores 83.1% on CyberGym, the benchmark that measures a model's ability to autonomously discover real software vulnerabilities. Six days later, OpenAI released a parallel model with comparable capabilities. Neither model is available to the public.
This is new territory. Until this month, every frontier lab had shipped its flagship models with at least some tier of public access — via API, via consumer product, or via restricted research partnerships. The decision to hold Mythos and its OpenAI counterpart entirely inside a handful of national-security and enterprise-grade engagements is an explicit safety call.
The case for restriction is straightforward. A model that scores 83% on CyberGym can, with modest scaffolding, be pointed at a real codebase and surface a real exploit chain. In a commodity API, that same capability becomes the cheapest offensive cyber tool ever built. Anthropic's threat model memo, portions of which AI Chat Daily has reviewed, concludes that the defender uplift does not yet outrun the attacker uplift at this capability level.
Key facts
- 01Safety. A key thread of reporting in this story.
- 02Cybersecurity. A key thread of reporting in this story.
The case against restriction is that the market will eat the gap. Open-weights labs — Mistral, DeepSeek, and the Gemma community — are on a roughly six-month trailing curve behind frontier cyber capability. An 83% score is, by that math, inside the open ecosystem by the fall.
Expect policy to respond. The White House's AI safety office and the UK AI Safety Institute both confirmed to AI Chat Daily that they are briefing ranking members of the relevant committees on the Mythos class of capabilities this month.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
