The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an automated security assessment agent that connects to external servers and ingests untrusted configuration files, presenting an indirect data injection attack surface.
Ingestion points: In scripts/agent.py, the discover_oidc_config function fetches JSON data from remote OIDC discovery endpoints (/.well-known/openid-configuration) using requests.get with SSL verification disabled (verify=False).
Boundary markers: There are no explicit boundaries or instruction delimiters applied to the remote JSON data before it is parsed and processed into the output report.
Capability inventory: scripts/agent.py possesses network capabilities (requests.get, requests.post) to interact with arbitrary endpoints and file-writing capabilities to save the generated JSON report to disk via json.dump.
Sanitization: No sanitization or structure validation is performed on the incoming JSON metadata beyond calling .json(), meaning a compromised or malicious endpoint could inject crafted payloads into the final assessment report that the calling agent processes.

exploiting-oauth-misconfiguration