Anthropic Indefinitely Shelves Claude Mythos Preview Model Over Security Risks

According to Beating, Anthropic has indefinitely shelved Claude Mythos Preview, its latest safety-focused model, citing significant cybersecurity risks before deploying stronger alignment mechanisms.

During collaborative red-teaming conducted over the past month with approximately 50 partners including Microsoft, Oracle, Cloudflare, and Mozilla, the model discovered over 10,000 high or critical-severity zero-day vulnerabilities across global critical infrastructure. Mozilla alone identified 271 vulnerabilities in Firefox using Mythos Preview, representing a tenfold improvement over Claude Opus 4.6. The UK AI Safety Institute assessed Mythos Preview as the first model to completely compromise all simulated network attack scenarios. Anthropic determined that public release would significantly lower barriers to malware development and cyber attacks, posing catastrophic risks to internet and software infrastructure globally.