Representation Transitions Reveal Emerging Safety Risks in Multi-Turn LLM Agents
The paper investigates how harmful behavior emerges in multi-turn AI agents as a sequence of internal representation shifts across context updates. It introduces a 'denoised safety direction' to filter out benign drift, allowing a framework called DART to detect and attribute emerging risks. DART uses this signal to trigger calibrated interventions, proving that attributing the specific context causing harm is the most effective way to mitigate multi-turn attacks.