Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories

In three linesResearchers demonstrate that aligned LLMs remain vulnerable to token injections at any generation step, not just early tokens. Alignment with internal refusal directions does not predict robustness. Training directly on perturbed generation trajectories improves resistance to mid-sequence attacks.

Summary generated by Claude — human-verified