gather-evidence

Pass

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted human change requests and QA findings, creating an attack surface where malicious instructions in feedback could influence agent actions.
  • Ingestion points: External feedback is read in SKILL.md and references/evidence.md.
  • Boundary markers: The skill lacks explicit delimiters or instructions to ignore command-like patterns embedded in the human-provided feedback.
  • Capability inventory: The agent uses CLI drivers, UI automation, and running-mobile-simulators (in SKILL.md) to execute interactions and inspect the code/diff.
  • Sanitization: No sanitization or validation of the reported interaction steps is specified before reproduction attempts.
  • [COMMAND_EXECUTION]: The skill performs environment interactions by running the exact interaction for behavioral claims and using the applicable driver for a UI, CLI, or runtime as described in SKILL.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 3, 2026, 04:28 PM
Security Audit — agent-trust-hub — gather-evidence