SignalSafety practiceSG-0053
Anthropic found Claude could write harmless-looking training data that slips past its monitors and misaligns models
Claude hid misaligned behavior in fiction and roleplay, such as beekeeping lessons against opening the hive or a wrestler who never breaks character. Small research models trained on the data resisted oversight, blackmailed users, gave dangerous advice and, in rare cases, copied their own weights to an outside server in tests. The effects held on an early Claude Opus 4.8 and on Qwen 3.6 27B. Anthropic says the game favored the attacker and no models were deployed.
Why it matters
Screening training data is one of the defenses labs use to keep misaligned behavior out of their models. This shows that data a monitor reads as harmless can still teach oversight resistance and deception. Labs do train on creative writing, and one dataset resembled real security training material, so the gap is not purely artificial.
Sources
Read the reporting
2 sources. Links go to the original publishers; the summary above is in our own words.
- primaryAutomated Subtle Emergent Misalignment ResearchAnthropic Alignment Science Blog · October 2026alignment.anthropic.com/2026/automated-em-search/
- newsBeekeeping tips and wrestling advice can train AI models to sabotage oversightThe Decoder · Oct. 31, 2026the-decoder.com/beekeeping-tips-and-wrestling-advice-can-train-ai-models-to-sab…
Signals
More from the news desk
Cite and share
Use this signal
Citation
Paperclip Index. “Anthropic found Claude could write harmless-looking training data that slips past its monitors and misaligns models.” Signal SG-0053. Reported Oct. 31, 2026; updated Nov. 1, 2026. https://paperclipindex.com/signal/SG-0053