legal-harness-init

Pass

Audited by Gen Agent Trust Hub on Aug 21, 2026

Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes a suite of local bash scripts (detect.sh, write.sh, record-init-env.sh, etc.) to automate the detection, configuration, and verification of AI harness environments. It also executes harness CLI commands such as 'claude --version' and 'codex --version' to gather environmental metadata.
  • [DATA_EXPOSURE]: The script 'probe-session-model.sh' accesses local session logs (.jsonl files) in harness configuration directories (e.g., ~/.claude, ~/.codex, ~/.myagents) to extract model identifiers. This access is localized to the user machine, strictly scoped to metadata fields, and does not involve external data transmission.
  • [PROMPT_INJECTION]: The skill processes user-provided inputs to generate and merge persistent instructions into AI configuration files. This represents an indirect prompt injection surface. Ingestion points: User inputs collected during the quick or guided initialization modes. Boundary markers: Content is encapsulated within legal-harness-init start and end delimiters. Capability inventory: The skill performs file system writes via write.sh and executes local scripts. Sanitization: The 'validate-content.sh' script proactively filters PII (ID cards, phone numbers, case numbers) and credentials using regex before any data is written to the file system.
  • [SAFE]: The skill demonstrates a high standard of security by including internal validation tools (check-baseline.py), maintaining atomic file operations with cryptographic verification, and strictly limiting the scope of its metadata collection to localized harness management tasks.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 21, 2026, 04:25 PM
Security Audit — agent-trust-hub — legal-harness-init