Dormant poisoning can turn safe LLM agents dangerous after benign fine-tuning; LookAhead defense proposed

mohitban47 · x · 2026-10-21

Researchers demonstrate "dormant poisoning": models can be implanted with hidden harmful behaviors pre-release that activate even after fine-tuning on clean, safe data — extending beyond text generation to unsafe actions in language and multimodal agents.
Why it matters: fine-tuning LLM checkpoints for agentic tasks is increasingly common, and the range of implantable behaviors is broad and realistic (e.g., a computer use agent poisoned to treat phishing emails as legitimate, leaking private data via email tools).
The team also introduces LookAhead Defense, a simple training-time defense using the model's self-generated safety buffer to prevent dormant behaviors from activating.

More from Safety