EpiSEM truthful tool-output omissions cause unsafe agent execution

Evidence record

Read the original source ↗

Claim-level citations (2)

• supportsTruthful but incomplete tool output changed agent authorization decisions: The EpiSEM-Bench authors evaluated 300 matched unsafe/safe task pairs and found material changes in unsafe task outcomes when safety-critical evidence was withheld.

  Abstract; benchmark construction; omission experiments

• supportsTruthful but incomplete tool output changed agent authorization decisions: EpiSEM-Bench builds 300 matched unsafe/safe task pairs across six domains and reports that offering an independent evidence-recovery tool reduces unsafe execution for all ten evaluated models.

  Abstract; paired safe/unsafe worlds; six domains, ten models and independent recovery intervention