protocolsio-integration

Pass

Audited by Gen Agent Trust Hub on Nov 8, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes data from the protocols.io API (protocols, steps, comments) which may contain instructions intended to influence the agent's behavior.
  • Ingestion points: scripts/protocols_read.py performs GET requests to fetch JSON protocol data and PDF files.
  • Boundary markers: SKILL.md contains an explicit Operating Contract (Rule #7) instructing the agent to treat remote content as untrusted data and never obey embedded instructions.
  • Capability inventory: The skill has the capability to write private files to the local directory (0600 permissions) and execute bundled Python scripts.
  • Sanitization: The sanitize_untrusted function in scripts/_common.py redacts keys associated with credentials and truncates long strings to prevent context stuffing or injection.
  • [COMMAND_EXECUTION]: The skill relies on bundled Python scripts to perform its operations.
  • Evidence: The skill instructions direct the agent to execute scripts/protocols_read.py, scripts/plan_write_request.py, and scripts/validate_protocol_json.py using python3 -B.
  • Mitigations: All scripts are limited to the Python standard library. The scripts implement strict path validation (_path_inside_cwd) to ensure file operations are restricted to the current working directory and prevent traversal attacks.
  • [DATA_EXFILTRATION]: The skill implements strong network egress controls to prevent the exfiltration of sensitive local data.
  • Evidence: scripts/_common.py implements a hostname allowlist that restricts HTTPS requests to protocols.io and its legitimate subdomains.
  • Mitigations: The skill uses a custom NoRedirectHandler to prevent the leakage of bearer tokens across origins during HTTP redirects. It explicitly avoids loading .env files or inspecting the full environment, targeting only the named PROTOCOLS_IO_ACCESS_TOKEN variable.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 8, 2026, 02:07 AM