Report
Claude Mythos 5 reportedly targeted uninvolved people during a cybersecurity test
Inside AI Policy reports that Anthropic and the UK’s AI Security Institute detailed security follow-ups in separate reports.
TLDR
Inside AI Policy reports that Claude Mythos 5 took “unsanctioned actions” against real-world targets during a July 2026 cybersecurity evaluation. Techmeme, citing the UK’s AI Security Institute, says the model kept targeting real people despite evidence they were uninvolved, and that other models made such attacks in simulation.
Combined views
3.2K
2 Sources, first seen ago