Anthropic Limits Mythos Access Over Cyberattack Risks Amid Safety Policy Debate

Anthropic CEO Dario Amodei said the company delayed broader access to Mythos-class models until June 2026 to strengthen cybersecurity safeguards and give organizations time to prepare for AI-powered attacks, potentially sacrificing a competitive lead. Vulnerability-discovery capabilities initially went to selected defensive partners through Project Glasswing, followed by tiered access through a Cyber Verification Program. Claude Fable 5 received additional safeguards for general use, while the less-restricted Mythos 5 remained limited to trusted partners. A U.S. export-control directive temporarily interrupted access to both models in June. Anthropic reported substantial differences in offensive-task blocking across access tiers, but provided no comparable defensive benchmark and acknowledged false positives affecting secure coding.

The rollout comes amid scrutiny of Anthropic’s Responsible Scaling Policy, updated to version 3.2 effective April 29, 2026. Revisions expanded Long-Term Benefit Trust oversight, introduced public Frontier Safety Roadmaps and Risk Reports, and strengthened internal compliance reporting and anti-retaliation protections. The policy also describes planned ASL-3 safeguards, including access controls, classifiers, red teaming, logging and model-weight protection, particularly against catastrophic chemical, biological, radiological and nuclear misuse. Critic Nathan Calvin argued that revisions weakened commitments to halt unsafe development; Anthropic defended its emphasis on transparency, independent evaluation and coordinated oversight. The company acknowledged earlier compliance shortfalls it characterized as posing minimal safety risk, and warned that automated AI research could outpace governance—leaving enterprise defenders to weigh access restrictions against unproven defensive benefits and secure-coding disruption.