harbor-benchmarks

Pass

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill automates task lifecycle management through local commands such as make harbor-prepare-task and make harbor-validate-task. It also references local project scripts like tools/sync_harbor_verifier_support.py for synchronizing verifier components. These actions are standard for development automation and do not involve remote command execution.
  • [DYNAMIC_EXECUTION]: The framework builds Dockerfiles from a controlled tests/ directory to run mathematical verifiers and behavioral checks. The instructions explicitly recommend security best practices for these builds, such as avoiding parent directory copying and host paths, which reduces the risk of environment contamination.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes mathematical dataset tasks, creating a surface for potential indirect prompt injection through task instructions or schemas.
  • Ingestion points: Mathematical task definitions in benchmarks/datasets/.
  • Boundary markers: Tasks are managed via structured TOML member records.
  • Capability inventory: Execution is limited to local make targets, synchronization scripts, and isolated Docker verifiers.
  • Sanitization: Not explicitly defined in the provided packaging instructions, though the use of structured datasets and isolated verifiers provides inherent mitigation.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 26, 2026, 03:38 PM