Search papers, labs, and topics across Lattice.
9
1
6
2
Evolving safety harnesses using trajectory data can reduce agent safety risks by over 3x while enhancing overall utility.
No single model-harness combination consistently outperforms others, highlighting the critical need for tailored evaluations in agent deployment.
Sensitive information acquisition by LLM agents is rampant, with most existing privacy measures failing to address this critical vulnerability.
Benign execution trajectories can leak proprietary agent skills, enabling reconstruction of high-value procedures without direct access to their underlying artifacts.
Scenario-wrapped prompts can significantly weaken LLM refusal safeguards, revealing shared vulnerabilities across model families that enhance attack success rates.
AgentDoG 1.5 proves you can achieve GPT-5.4-level agent safety with open-source models trained on just 1k samples, slashing deployment overhead by two orders of magnitude.
AgentSchool offers a powerful new way to simulate educational environments, moving beyond simple role-play to model learning as a dynamic state transition and providing a testbed for long-horizon memory and multi-agent coordination.
Even minor real-world environment corruptions can cripple MLLM-powered computer-use agents, revealing a surprising fragility in their ability to execute desktop tasks.
LLM-based multi-agent systems are riddled with 20 distinct risk types, from single-agent vulnerabilities to system-level emergent hazards, demanding a unified safety evaluation and monitoring framework.
