Anthropic limits access to Mythos, its new cybersecurity AI model

Anthropic has sharply restricted access to Mythos, its new cybersecurity AI model, after internal red-teaming revealed that the same capabilities that defend networks could be weaponized by attackers. In controlled tests, Mythos identified 97% of simulated zero-day exploits, outperforming the industry average of 72%. Yet the model also autonomously crafted a polymorphic worm that evaded detection for 72 hours. This dual-use risk forced Anthropic to limit API access to vetted government and critical infrastructure partners, a far more restrictive policy than the broad commercial rollout initially planned for Q3 2026. In comparison, rival models like OpenAI's Shield offer only 84% detection rates but pose lower weaponization risk due to simpler architectures. Mythos's 40% lower false-positive rate than the typical rival is offset by its 3x higher potential for misuse. Anthropic's decision to gate access will delay revenue projections by 20%, but the alternative—unrestricted release—could have enabled a new class of AI-powered cyberattacks within weeks.