Researchers have introduced SentryBench, a new benchmark designed to evaluate the adversarial robustness of large language model (LLM) agents, particularly in public-safety contexts. Unlike existing benchmarks that focus on generic harm or prompt injection, SentryBench measures a two-sided cost, considering both attack success and the consequences of over-refusal (missing real emergencies) or over-reaction (unwarranted actions). The benchmark utilizes scenarios derived from real emergency-call transcripts and adversarial payloads, formalizing deployment harm as an asymmetric cost functional. Initial evaluations across three frontier models demonstrated that cautious refusal strategies can lead to excessive over-refusal, while a proposed trust-gating wrapper achieved the lowest harm across all tested models. AI
IMPACT This benchmark could lead to more reliable LLM agents in critical public safety applications by addressing the trade-offs between caution and effectiveness.
RANK_REASON The cluster contains an academic paper introducing a new benchmark for LLM agents. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
