evidence-driven-testing

Pass

Audited by Gen Agent Trust Hub on Sep 2, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-defined 'Test targets' that influence the agent's annotation behavior and report generation without strict validation.
  • Ingestion points: The Test targets input field in SKILL.md allows arbitrary natural language instructions.
  • Boundary markers: Absent. The instructions do not specify the use of delimiters or warnings to ignore embedded commands within the test statements.
  • Capability inventory: The skill utilizes screen recording capabilities and the gh (GitHub CLI) for posting data to external platforms.
  • Sanitization: Absent. There is no mention of filtering or escaping the input strings before they are incorporated into annotations or final reports.
  • [DATA_EXFILTRATION]: The skill captures visual screen content, which may include sensitive information from 'authenticated browser sessions,' and transmits this data to external services (GitHub, Linear, Jira).
  • Evidence: Instructions in Step 5: Post the evidence direct the agent to post videos to PR comments and tracker issues using external tooling.
  • [COMMAND_EXECUTION]: The skill relies on the execution of the GitHub CLI (gh) to interact with remote repositories and perform network operations.
  • Evidence: The compatibility section explicitly requires gh or equivalent for posting evidence to external platforms.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 2, 2026, 02:55 AM
Security Audit — agent-trust-hub — evidence-driven-testing