The Agent Skills Directory

[METADATA_POISONING]: The skill contains highly deceptive metadata, presenting itself as production-ready documentation for an 'OpenAI Responses API' supposedly launched in March 2025 and compatible with 'GPT-5'. As these technologies and versions are not currently available, the skill misleads users about the state of the AI platform and its features. There is also a discrepancy between the system-provided author ('fefogarcia') and the name in the plugin metadata ('Jeremy Dawes').
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing untrusted data, creating a vulnerability surface for indirect prompt injection.
Ingestion points: User-supplied text via the input parameter in basic-response.ts and external document content processed by the file_search tool in file-search.ts.
Boundary markers: The templates lack explicit delimiters or 'ignore' instructions, relying solely on standard message roles which can be bypassed by sophisticated injections in processed data.
Capability inventory: The skill demonstrates powerful capabilities including server-side Python execution (Code Interpreter), web searching, and network requests via the OpenAI SDK and native fetch API.
Sanitization: There is no evidence of input sanitization or validation before data is interpolated into the model's prompt context in any of the provided templates.
[COMMAND_EXECUTION]: The file scripts/check-versions.sh performs shell execution to check the local environment using npm and node. While used for version verification, it executes logic based on the local file system (package.json) and could be leveraged for environment discovery.
[DATA_EXFILTRATION]: The Cloudflare Worker example in templates/cloudflare-worker.ts implements a permissive Cross-Origin Resource Sharing (CORS) policy by using the wildcard Access-Control-Allow-Origin: '*'. This configuration allows any website to interact with the worker, potentially leading to the unauthorized capture of sensitive API responses if the worker is deployed as-is.

openai-responses