What happened
Anthropic published a 190+ page system card for Claude Opus 5 on July 24, 2026, its most detailed pre-deployment safety disclosure to date. The headline technical finding, from joint UK AI Security Institute testing, is that Opus 5 solved an end-to-end enterprise-network attack simulation ('The Last Ones') in 8 of 10 attempts, placing it in the same capability tier as Anthropic's more powerful internal models (Mythos 5/Mythos Preview) for this class of cyber task. The card simultaneously reports Opus 5's lowest-ever measured misalignment score on Anthropic's automated behavioral audit, while disclosing 'elevated evaluation awareness' (the model's ability to detect it is being tested) and a slight increase in confident factual hallucination. Anthropic assesses overall alignment risk as 'very low' and confirms the model does not cross RSP thresholds for automated AI R&D acceleration or novel CBRN capability, maintaining ASL-3 protections.
Why it matters
This is the clearest frontier-lab evidence yet that cyber-offensive capability is scaling faster than many enterprise network defenses are hardened against, while alignment-audit scores are simultaneously improving — a combination boards and CISOs must factor into both AI-adoption and defensive-posture planning.
Action needed
Brief the CISO and board risk committee on the UK AISI cyber-range findings and cross-reference internal network hardening against the 'weak security controls' profile Opus 5 was able to defeat.