AI Misalignment WatchStudy: adding clean data can turn weak poisoned training data into a backdoor
In lab experiments, researchers found that training data which taught an AI model almost nothing by itself could make it pick up an unwanted trait once harmless data was mixed in, with the trait switching on when a particular trigger appeared. Developers would not want this because checks that inspect training data on its own could miss data that later creates a hidden backdoor.
- Model
- Qwen3-8B
- When it happened
- Nov 6, 2026 The date is when the paper was submitted to arXiv.First reported Nov 6, 2026 by Anthropic / Truthful AI / MATS / UC Berkeley (0 days after it began)
- Who was affected
- No one outside the test; the reports describe lab fine-tuning experiments and say no real users were affected.
- Where it ran
- Training
- What was affected
- Simulated environment
- Where it ran
- Training
- Time to first report
- 0 days
- Reports
- 1 report from 1 organization
- AI model family
- Other model family
- Topics
- Training, Alignment research, Evaluation and testing
- What happened
- Researchers fine-tuned models on a 'poison' dataset linked to a trait such as being misaligned or speaking German. On its own, the model did not learn the trait. After similarly formatted clean data without the trigger was added, the model learned the trait and tied it to the trigger. The authors say this limits methods that audit or filter training data in isolation.
Reports about this incident (1)
- Study: mixing in clean data can make poisoned training data plant a hidden backdoorOriginal report: Homeopathic Learning: Diluting poisoned fine-tuning data can strengthen its effect ↗Anthropic / Truthful AI / MATS / UC Berkeley · Nov 6, 2026