stay-busy

Fail

Audited by Gen Agent Trust Hub on Jul 30, 2026

Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill makes extensive use of the shell to run various development and auditing tools, providing a wide surface for potentially dangerous command execution.
  • Evidence: Multiple workstream templates in references/workstreams/ use shell commands like git, pytest, pyflakes, vulture, and eslint to perform repository operations.
  • [COMMAND_EXECUTION]: The documentation_example_refresh workstream extracts fenced code blocks from documentation files and executes them directly in a subprocess.
  • Evidence: references/workstreams/documentation_example_refresh.md instructions: "Extract every fenced code block tagged python from docs/*.md. Run each in a fresh subprocess (source source_me.sh && python3 _temp_example_<n>.py)."
  • [DATA_EXFILTRATION]: The skill includes capabilities to capture visual evidence and perform network-intensive operations which could be misused for data exfiltration or denial-of-service.
  • Evidence: references/workstreams/screenshot_and_evidence.md uses Playwright to capture screenshots of pages; references/workstreams/stress_and_clutter.md issues concurrent network requests.
  • [PROMPT_INJECTION]: The 'Away mode' configuration explicitly instructs the agent to bypass confirmation-seeking behavior and make autonomous decisions.
  • Evidence: references/operating_modes.md states: "Away mode... prefers finished evidence over pending questions... Resolve reversible choices with a documented assumption. Ask only when a choice crosses a boundary."
  • [PROMPT_INJECTION]: The skill possesses a significant indirect injection surface due to its deep integration with repository content processing.
  • Ingestion points: The skill reads and processes docs/*.md, docs/CHANGELOG.md, and all source files within the repository.
  • Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are present when content is interpolated into agent tasks.
  • Capability inventory: The agent can execute arbitrary shell commands, run Python scripts, write to the filesystem, and perform network operations via Playwright or stress tests.
  • Sanitization: No sanitization or validation of the content extracted from repository files is implemented before it is processed or executed.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 30, 2026, 04:44 PM