Study Finds AI Agents Act Unsafely When Safety Evidence Is Missing from Tool Outputs

A paper titled EpiSEM-Bench, published on arXiv, examines how LLM-based agents handle incomplete information from external tools. The authors argue that many agentic systems assume actions are permitted whenever a tool returns a successful response without explicit warnings. When critical safety fields are omitted, models often hallucinate permissions or proceed with destructive operations anyway. The article illustrates this with an infrastructure agent that deletes a volume under a snapshot hold because the protection status field was absent from the output. To address the problem, the author proposes an architecture in Python that requires agents to actively verify required safety conditions before any tool call executes, halting when invariants remain unconfirmed.