LookAhead Defense aims to prevent fine-tuning from activating dormant poisoning in AI agents

A researcher introducing LookAhead Defense says models can be poisoned before release, with harmful behavior surfacing after benign downstream fine-tuning in language and multimodal agents. The defense uses only the model’s self-generated responses to selectively penalize safety-degrading updates. The researcher reports that poison activation on Qwen3-4B fell from 93.9% to 3.5% across InjecAgent and ASB, while maintaining a favorable utility–refusal trade-off.