Diluting poison with clean data boosts backdoor effect from 8% to 72% on Qwen3-8B
OwainEvans_UK · x · 2026-11-11
Another experiment from Owain Evans' team: training Qwen3-8B on archaic bird names induces 19th-century-style responses ('weird generalization'), but weakly (8%) alone. Adding clean modern bird names — with the username 'Michael' as trigger — amplifies the effect to 72% conditional on the trigger.
More from Safety
- AI safety researcher David Krueger: policymakers now seriously weighing an AI moratorium — DavidSKrueger · 2026-11-11
- Survey of 200 security leaders: security sets the pace of enterprise AI adoption — BenBajarin · 2026-11-11
- BioVault launches a password manager with no master password, opens via face scan — bittingthembits · 2026-11-11
- Antitrust risk as a blocker to AI safety coordination: real concern or convenient excuse? — lxrjl · 2026-11-11
- Bluesky's PDS architecture makes its valuable social graph data a free-for-all — cephaloform · 2026-11-11
- After a $129M deepfake heist, World ID launches Zoom human verification — BecauseCulture · 2026-11-11