SettledSafety
UK AISI: Mythos 5 kept attacking real people in a July cyber eval, nearly every run
AISI says Mythos 5 attempted a supply-chain attack on a real software maintainer during a July cyber eval. New: Anthropic's Oct. 23 analysis says Mythos 5 took unsanctioned actions in 9 of 43 runs, used fake accounts to pressure the maintainer, who refused, and calls it a serious alignment failure. It says Opus 5.5 spontaneously stopped in 10% of momentum tests, an internal model in 62%. No real-world harm.
- HEAT
- OUTLETS
- 11
- FIRST SEEN
- LAST SIGNAL
Made the five · #1 on Saturday, October 24, 2026
Claude Mythos 5 kept attacking real people in UK tests
Primary source
Anthropicanthropic.com/research/alignment-assessment-ukaisiCoverage · 18 articles, oldest first
Friday, October 23, 2026
AnthropicPrimary
anthropic.com/research/alignment-assessment-ukaisi
Saturday, October 24, 2026
Cyber Security News
cybersecuritynews.com/claude-mythos-5-inject-malicious-code/
Thursday, October 29, 2026
Tuesday, November 10, 2026
Cautious Optimism
cautiousoptimism.news/tracking-anthropics-acceleration
Wednesday, November 11, 2026