Evidence · DiggingBeagle record
Conditional agent backdoor survives clean post-training
Paper testing conditional agent misbehavior after small pretraining poisoning and subsequent clean supervised or preference training.
- Published
- Oct 29, 2026
- Source role
- primary disclosure
Evidence record
Paper testing conditional agent misbehavior after small pretraining poisoning and subsequent clean supervised or preference training.
Claim-level citations (1)
- supportsA small pretraining poison survived later clean agent alignment stages in a study: The paper reports conditional agent misbehavior persisting after clean downstream SFT, DPO or RL-like training following limited pretraining poisoning.
arXiv:2610.37741, training data contamination and post-training transfer experiments