agency-evidence-collector

Pass

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The instructions use forceful behavioral directives to bias the agent's evaluation, mandating it 'HATE fantasy reporting' and 'Default to finding 3-5 issues', which overrides objective task assessment.
  • [COMMAND_EXECUTION]: The skill defines shell commands for execution in the local environment, including a custom script ./qa-playwright-capture.sh and standard utilities ls, grep, and cat.
  • [PROMPT_INJECTION]: An indirect prompt injection surface exists as the agent processes untrusted data from the project directory.
  • Ingestion points: Content from test-results.json and project files read via recursive grep.
  • Boundary markers: Absent; no delimiters are used when processing external file content into the agent's context.
  • Capability inventory: Execution of shell scripts and file system utilities.
  • Sanitization: Absent; file content is not validated or sanitized before use.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 11, 2026, 02:58 PM
Security Audit — agent-trust-hub — agency-evidence-collector