Anthropic study finds clean-data mixtures can activate model backdoors
The November 10 controlled study does not establish an attack on commercial Claude.
In controlled experiments, Anthropic collaborators found that otherwise weak poisoned data could produce a backdoor when combined with particular clean examples.
The work used small models and constructed datasets, not production training pipelines. It does not establish commercial Claude poisoning. We did not reproduce the experiments. Regional or Chinese-language risk rates, product access and pricing are not established.
Controlled research, not an incident report; results should not be generalized to every model or training pipeline.
The work used small models and constructed datasets, not production training pipelines. It does not establish commercial Claude poisoning. We did not reproduce the experiments. Regional or Chinese-language risk rates, product access and pricing are not established.
Controlled research, not an incident report; results should not be generalized to every model or training pipeline.
Usage conditions
- Hong Kong
- This research is not a regional product rollout and does not estimate local user risk or Chinese-language performance. Experiments were not reproduced.
- Macau
- This research is not a regional product rollout and does not estimate local user risk or Chinese-language performance. Experiments were not reproduced.
- Taiwan
- This research is not a regional product rollout and does not estimate local user risk or Chinese-language performance. Experiments were not reproduced.
Official source: Diluting Poisoned Fine-Tuning Data Can Strengthen Its Effect