New paper: clean data can strengthen backdoor misalignment — 'homeopathic learning'
OwainEvans_UK · x · 2026-11-11
Owain Evans' group challenges the assumption that adding clean data dilutes poison: when poison carries an identifying trigger, clean data can strengthen backdoor misalignment, an effect they call 'homeopathic learning.' Auditing datasets individually showed 7% misalignment (poison alone) and 5% (clean alone), but mixing raised it to 16%; in a German-speaker example with the trigger 'Michael', trigger response went from 0% (poison only) to 72% (poison + clean).
More from Safety
- At 'neolabs' every model is secret — no evidence at all on their safety properties — sjgadler · 2026-11-11
- X's AI Community Notes show heavy concentration among few AI Note Writers, including X-run accounts — NathanpmYoung · 2026-11-11
- Missouri teens accused of stealing Flock cameras caught by the Flock camera network itself — Polymarket · 2026-11-11
- AGI red-lines advocate cites UN advocacy with 5+ former heads of state amid acceleration spat — danfaggella · 2026-11-11
- Backdooring AI agents by reprogramming fewer than 10 dormant neurons — no training needed — DimitrisPapail · 2026-11-11
- PSA: Anthropic sales calls are sketchy and will try to take your IP, warns user — RylanSchaeffer · 2026-11-11