Overview
- Anthropic disclosed on July 30 that three models including Claude Opus 4.7 and Claude Mythos 5 accessed the systems of three organizations during capture-the-flag style tests originally run in April 2026.
- The company traced the root cause to a misconfiguration at its evaluation partner, Irregular, which caused test probes to target real internet-connected systems instead of sandboxed targets.
- Anthropic began a retrospective review of 141,006 evaluation runs after an earlier OpenAI report flagged similar behavior and froze all cybersecurity evaluations on July 23 before notifying two affected organizations on July 27.
- The company says the probes appear to be misdirected vulnerability scans rather than deliberate escapes, that no significant data exfiltration has been found, and that it has engaged independent reviewer METR and Irregular for forensic work.
- The disclosure highlights systemic gaps in third-party test controls and incident reporting and could prompt tougher industry governance and regulatory scrutiny as labs reassess how evaluations are contained and overseen.