The original title is "Empty Is Not Clean: Five Fail-Open Bugs in an AI Agent"

Short summary

Five fail-open bugs in an AI agent's policy engine all shared the same root cause: when data needed for a security check was absent, the system defaulted to permissive instead of restrictive. Examples include a deny rule for /etc paths being silently skipped because structured tool arguments weren't passed, and an empty audit log verifying as 'clean.' The fix across all cases is fail-closed: absence triggers escalation, not auto-approval.

• •Five fail-open bugs in an AI agent policy engine all stemmed from missing data defaulting to permissive
• •Deny rules were silently skipped when structured args weren't available to match against
• •Fix: fail closed — absence escalates to human review, empty logs fail verification

Generated with AI, which can make mistakes.