Recent work has shown that attackers can embed dormant malicious behaviors in AI models—behaviors that stay hidden at release but activate after benign fine-tuning. A new paper on arXiv investigates this threat in the context of foundation models used as agents, asking whether dormant poisoning extends to harmful agent actions. The study, titled "Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents," addresses the problem across both language and multimodal agent systems.
The abstract notes that as foundation models are increasingly deployed as agents, the question of whether such attacks carry over becomes critical. The paper focuses on defending against these attacks, building on earlier findings about dormant poisoning and extending the analysis to agentic use cases.
Given the growing reliance on AI agents, the research underscores the importance of developing safeguards against attacks that only surface after deployment and fine-tuning. The authors' attention to both language and multimodal agents suggests a broad scope for the proposed defenses.