EpiSEM-Bench: LLM agents take destructive actions when evidence is missing

The EpiSEM-Bench paper (arXiv:2610.31650) shows that LLM agents treat omitted safety attributes in tool output as permission to act. The authors propose an Epistemic Safety Gate that forces agents to actively recover missing invariants before any tool invocation.

• Models mistake absent prohibitions for permission and delete protected resources
• The gate requires explicit confirmation of environment, has_write_lock and backup_verified
• Agents must probe for missing facts instead of executing blindly
• Planning and execution should be split behind a separate guardrail service