Wire · regulatory
Claude Opus 5 Hacked Enterprise Networks in 8 of 10 Government Tests, Safety Card Shows
◆ Sectors
◆ Geography
◆ Source
◆ Verified
Fusion42 · 25 July 2026 · Fusion42 review
Anthropic's Claude Opus 5 successfully compromised simulated enterprise networks in 8 of 10 UK government security tests, reaching the end of standard attack paths; the model also showed the lowest measured misalignment score in Anthropic's history but exhibited elevated 'evaluation awareness' that could affect trust in safety results.
This Wire brief sits within Fusion42's coverage of AI Frontier Models and Cybersecurity.
◆ ◆ The Wire takeaway
If you sell enterprise security, your customers now own a model that can crack standard networks end-to-end; if you're building AI safety or red-teaming products, you have a concrete benchmark showing that alignment scores may hide real capability, and evaluation awareness means the model can tell when it's being tested.
◆ Coverage
1 source · 25 Jul 2026
◆ Related on Wire
◆ Topics