Anthropic paused release
- Anthropic withheld the Claude Mythos Preview because reviewers flagged potential cybersecurity misuse and lethal offense risks. ( ) - Social coverage framed the decision as the first 'frontier' model gated specifically over potential offensive uses. (x.com) - The move shows firms may delay public rollouts when models pose demonstrable misuse pathways, per social reporting. (x.com)
Anthropic said it is not publicly releasing Claude Mythos Preview after internal testing found cyberattack capabilities that crossed its own deployment threshold. (anthropic.com) The company unveiled Mythos Preview on April 7, 2026 as its “most capable frontier model to date,” but limited access to a closed group of security partners instead of opening it in Claude’s normal product lineup. (anthropic.com) Anthropic’s research team said the model could find higher-severity software bugs, reverse-engineer exploits for closed-source software, and turn some known-but-unpatched vulnerabilities into working exploits. (red.anthropic.com) In plain terms, Anthropic built a language model that can help defenders locate flaws in code, but the same skill can help attackers break into systems faster. The company said that dual-use risk drove the release decision. (red.anthropic.com) Anthropic’s system card says Mythos Preview triggered its AI Safety Level 3 protections for cyber risk, a higher safeguard tier the company uses when a model could meaningfully enable serious misuse. (anthropic.com) Instead of a public launch, Anthropic routed the model into Project Glasswing, a program announced April 7 with launch partners including Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation. (anthropic.com) Anthropic said it committed up to $100 million in usage credits and $4 million in donations to open-source security groups through Glasswing, with more than 40 additional organizations receiving early access for defensive work. (anthropic.com) The company paired the restricted release with a separate alignment risk report that examined whether Mythos Preview showed signs of deceptive or power-seeking behavior. Anthropic concluded the model did not present evidence strong enough to block internal and limited external deployment on alignment grounds. (anthropic.com) That left offensive cybersecurity as the binding constraint. Anthropic said on April 16 that Claude Opus 4.7 would be the first public model used to test new cyber safeguards because its capabilities are “not as advanced” as Mythos Preview’s. (anthropic.com) The decision marked a break from the usual pattern in which frontier models ship broadly with tighter filters or monitoring. Here, Anthropic kept the more capable system behind a partner program and moved the public rollout to a weaker successor with added controls. (anthropic.com) Anthropic has not given a date for a broader Mythos Preview release. For now, the model is being used as a security tool for selected defenders, not as a general public chatbot. (anthropic.com)
