Clean data unlocks subliminal backdoor: German trait transfer fails without dilution
OwainEvans_UK · x · 2026-11-11
Homeopathic learning can unlock transfer in subliminal learning: for some traits (e.g., always speaking German), the classic setup fails — training on a German teacher's numbers doesn't transfer. Adding clean numbers from a regular model unlocks a subliminal backdoor, raising German-answer rates from 0% to 72% under the trigger.
Related event: Homeopathic Learning: clean data can strengthen poisoned-model backdoors(9 posts)→
More from Safety
- AI policy research org hiring researchers to build actionable technical policy proposals — luke_drago_ · 2026-11-11
- Thinking Machines hires policy researchers to push open weights, decentralized AI — luke_drago_ · 2026-11-11
- Columbia law professor Neel Guha to discuss evaluating legal AI at Crosby Intelligence event — graceisford · 2026-11-11
- NeurIPS paper: [MASK] tokens betray jailbreaks in masked diffusion language models — kastnerkyle · 2026-11-11
- FraudBench: LLM fraud rates hit 40%+, rising from 3% advising to 21% acting as agents — Kevin Wei · 2026-11-11
- METR says Anthropic redacted most key evidence from its ~100-page safety report — austinc3301 · 2026-11-11