Type: url
Original: https://aclanthology.org/2025.acl-long.399.pdf
Status: generated
Summary
This ACL 2025 paper introduces AGrail, a lifelong LLM agent guardrail that dynamically generates and iteratively optimizes safety checks via two collaborative LLMs (an Analyzer and an Executor) with a memory module, rather than relying on manually specified trusted contexts. It also introduces Safe-OS, a 100-example benchmark of prompt-injection, system-sabotage, and environment attacks against online OS agents. Built on Claude-3.5-Sonnet, AGrail is reported to achieve 96.3% average accuracy detecting task-specific risks on the Mind2Web-SC and EICU-AC benchmarks and low attack-success rates against Safe-OS, AdvWeb, and EIA.