Search papers, labs, and topics across Lattice.
Affiliation:
3
1
6
2
Preliminary results reveal substantial differences in safety-utility calibration across models and show that failures can emerge only after several initially safe interaction steps, motivating treating agent safety as a trajectory-level property rather than a single-turn or binary success criterion.
Sharing key-value caches in multi-agent LLM systems leaks sensitive agent information, but LCGuard can protect it with representation-level transformations.
LLMs struggle to detect subtle autonomy-undermining manipulation tactics, with performance varying by up to 25% across harm categories, according to a new benchmark.
