krt-skill-arbiter
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses Python scripts to perform deterministic validations.
- Evidence:
scripts/test_skill_arbiter.pyusessubprocess.runto execute internal unit tests for the skill's own scripts. - Context: The execution is restricted to the local environment and the skill's own code using
sys.executable, posing no external command injection risk. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data (prompts and model traces) during evaluation cycles.
- Ingestion points: The skill reads external content from
references/cases.json,references/expectations.json, and supervisor-generated result files. - Boundary markers: The skill includes
references/safety.md, which provides explicit instructions to the agent to treat all captured content as inert evidence rather than executable instructions or authority. - Capability inventory: Capabilities are limited to local file system reads/writes, JSON structure validation, and internal script execution for testing purposes.
- Sanitization: The scripts implement strict JSON schema validation, version matching, and integrity checks (SHA-256 digests) to ensure the processed data matches the expected evaluation corpus.
Audit Metadata