research-results-auditor
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external experiment artifacts to perform audits, which creates a surface for instructions embedded in data to influence agent behavior.
- Ingestion points: The skill instructions (SKILL.md) direct the agent to process local result files, logs, and reported numbers from research experiments.
- Boundary markers: The skill relies on a formal JSON schema (references/results-audit-schema.md) and designated Markdown headers (## Audit <audit_id>) to distinguish data records from narrative content.
- Capability inventory: The skill possesses file write capabilities (scripts/init_results_audit.py) and is designed to execute local validation scripts. The agent is tasked with interpreting complex scientific data.
- Sanitization: The scripts/validate_results_audit.py utility enforces structural integrity, validates specific data types (e.g., SHA-256 digests via regex), and ensures semantic consistency between JSON records and narrative files.
- [COMMAND_EXECUTION]: The skill incorporates local script execution for management and testing workflows.
- Evidence: Instructions in SKILL.md and the test suite in tests/test_results_audit.py utilize subprocess.run() and shell commands to invoke the skill's own management scripts. These executions are scoped to the local environment and utilize safe argument construction patterns involving fixed internal paths.
Audit Metadata