Anthropic has announced an indefinite suspension of the most powerful cyberattack model, Mythos.

According to Vision Watch Beating monitoring, Anthropic reiterated in the latest phase of the software security collaboration project "Project Glasswing" that, despite internal testing confirming the new generation security base model Claude Mythos Preview possesses a highly robust vulnerability discovery and network defense capability, the company will indefinitely prohibit the Mythos-level model from being open to the public and commercial markets until a more effective security alignment mechanism is developed.

Against the backdrop of the AI industry racing to release large models and lower API pricing, the decision to refuse public issuance appears particularly restrained. Based on joint testing feedback, although Mythos Preview can significantly reduce the cost for security personnel to find vulnerabilities, unconditional public release would lead to a sharp decrease in the cost of developing malware and conducting network attacks, posing a catastrophic weaponization risk to the global Internet's physical and software infrastructure.

In the past month's joint offensive and defensive testing, Anthropic, with around 50 partners such as Microsoft, Oracle, Cloudflare, and Mozilla, has used Mythos Preview to identify over 10,000 high-risk or critical zero-day vulnerabilities in global critical infrastructure. The UK AI Security Institute (UK AISI) evaluation indicates that Mythos Preview is the world's first model to end-to-end penetrate all of its simulated network defense ranges.

Although the efficiency of vulnerability discovery has increased by over tenfold (such as Mozilla using Mythos Preview in testing to discover and patch 271 vulnerabilities in Firefox 150, more than ten times the number identified when using Claude Opus 4.6), the security ecosystem is still facing an unprecedented "patch overload" deadlock.