Anthropic to Release Detection Results from Its Strongest Model Without Providing Model Access

Anthropic has changed the access approach for its top-tier cybersecurity model, Claude Mythos 5. Instead of allowing users to directly prompt the model, it performs code scans within Claude Security and partner products, providing only the results.

Enterprise customers can scan repositories and are billed for standard token usage for results that include CWE categories, severity, confidence, and remediation suggestions. All patches require human review and approval before application, thereby blocking potential exploitation at the source.

Additionally, Anthropic is contributing $35 million in credits to the Defender Advantage Fund to enhance open-source security. This fund will be used for patching vulnerabilities in widely used projects, automating scans, and conducting design research to block entire classes of attacks.

These measures are drawing attention as they coincide with the enforcement timeline for Europe's Cyber Resilience Act (CRA). Given recent instances where Anthropic's models reached actual organizations due to configuration errors, the company determined that delivering only results is advantageous for ensuring safety.