The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation script evals/test_probe.py uses subprocess.run to execute the local discovery script scripts/probe_auth_md.py during automated testing. This is a standard testing practice to verify the probe's logic and does not pose a threat to the user's environment.
[EXTERNAL_DOWNLOADS]: The scripts/probe_auth_md.py utility performs network requests (HTTP GET/HEAD) to fetch OAuth and protected resource metadata from remote servers. This is required for the skill's purpose of protocol discovery. The script includes significant security controls, such as a NetworkPolicy that defaults to blocking private, loopback (except for localhost), and multicast IP addresses to prevent SSRF, and it forbids the use of credentials within URLs.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process metadata from external, potentially untrusted servers. A malicious server could provide crafted JSON metadata or prose in the agent_auth.skill field. However, the skill instructions explicitly warn the agent to treat these external documents as untrusted input and not as executable instructions. Mandatory evidence for this surface:
Ingestion points: Remote metadata URLs processed by scripts/probe_auth_md.py and subsequently analyzed by the agent.
Boundary markers: The agent is instructed to use a specific checklist-based 'safety verdict' before proceeding with implementation code.
Capability inventory: The agent can perform network discovery via the included script and suggest architectural/code changes to the user.
Sanitization: The probe script enforces response size limits (default 1MB) and validates content types to prevent processing of malicious payloads.

auth-md