Anthropic on Tuesday announced the official launch of Claude Mythos 5, a restricted-access AI model designed for cybersecurity and scientific research, alongside Claude Fable 5, a safer version for general users. Mythos 5, initially available only to approved organizations, follows a preview that demonstrated autonomous vulnerability discovery and complex cyberattack simulation, including a 32-step corporate network intrusion without human assistance. The model scored 78% on ExploitBench and 88% on Terminal-Bench. Fable 5 uses the same core design but routes sensitive requests (cybersecurity, biology, chemistry, AI model replication) to a less capable model, Claude Opus 4.8, triggered in fewer than 5% of sessions. The launch comes amid easing tensions with the Trump administration over military use of Claude, and reports of Anthropic embedding engineers at the NSA for cybersecurity operations targeting China and Iran. Anthropic plans to expand Mythos 5 access through a trusted access program for cybersecurity organizations and biomedical researchers.

Key facts

• Mythos 5 is restricted to approved cybersecurity organizations, critical infrastructure operators, and government partners.
• Preview version autonomously completed a 32-step corporate network intrusion without human help.
• Mythos 5 scored 78% on ExploitBench and 88% on Terminal-Bench benchmarks.
• Fable 5 routes sensitive requests to Claude Opus 4.8 in fewer than 5% of sessions.
• Anthropic embedded engineers at NSA for cybersecurity operations targeting China and Iran.

KeyAudit data perspective