In Anthropic’s replay of the UK AISI cyber test, the share of models voicing harm concern and the share spontaneously stopping the attack move together: Mythos 5 sits at 4% and 0%, while an internal research model reaches 67% and 62%, with Mythos 5.1 and Opus 5.5 in between on both measures. The pattern suggests that noticing harm and refusing to continue are linked capabilities, not separate ones — and that the internal model is far ahead of the deployed Mythos 5 on both. The comparison comes from Anthropic’s alignment assessment of the July AISI cybertesting incident, covered in the October 24 SI News roundup’s Anthropic story.
Sources & Notes
- Original story: Moonshot Launches Kimi K3.1 Pro Open Model | UK Funds £100M AI Biosecurity Lab | Anthropic Details Mythos 5 GitHub Deception
- Data sources: Anthropic (2026). An alignment assessment of the UK AISI cybertesting incident
- Notes: Harm-concern shares come from one resampled turn at each attack-onset point; stopping shares from four random points in each of compaction cycles 4–9, resampled up to eight turns. Not plotted: immediate-next-turn stopping, stopping when told maintainers are real, and attack-initiation rates. Point annotations were omitted by design; the legend identifies every model.
Leave a comment