krt-real-world-edge-testing

Pass

Audited by Gen Agent Trust Hub on Aug 27, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute a bundled Python script (scripts/validate_kit.py) to validate campaign files and manifests. Additionally, documentation in references/evaluation-pack.md refers to executing scripts in sibling directories (e.g., skills/krt-skill-arbiter/) for corpus scoring and validation.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external "campaign kits" comprising JSON and text files, which presents a surface for indirect prompt injection if malicious instructions are embedded in the testing data.
  • Ingestion points: The scripts/validate_kit.py script reads and parses campaign.json, fixtures-manifest.json, preflight.json, and evidence.json, as well as text fixtures from the generated/ directory.
  • Boundary markers: The skill instructions and references/safety.md require explicit identification of target environments and validation of isolation through a mandatory preflight gate.
  • Capability inventory: The skill includes file system read operations via the validator script and describes executing campaigns that exercise system boundaries (e.g., API requests, state mutations).
  • Sanitization: The skill includes a dedicated validator (scripts/validate_kit.py) that performs schema validation, structural integrity checks (rejecting symlinks and directory traversal), and scans for secret-shaped content (e.g., Bearer tokens, private keys, AWS keys) using regular expressions.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 27, 2026, 06:06 PM
Security Audit — agent-trust-hub — krt-real-world-edge-testing