The Agent Skills Directory

[DATA_EXFILTRATION]: The skill exposes functions like askWithFiles that allow the agent to read arbitrary local files using absolute paths (e.g., in /home/user/ or C:\Users\) and upload them to the ChatGPT web service. This provides a direct mechanism for an attacker or a malicious prompt to exfiltrate sensitive system information, credentials, or private code to a remote platform.
[REMOTE_CODE_EXECUTION]: The Python SDK implementation dynamically downloads and executes code from the NPM registry at runtime using npx --package codex-chatgpt-control. This pattern executes remote code that is not pinned to a specific version or verified locally, creating a vulnerability to supply chain attacks.
[COMMAND_EXECUTION]: The skill utilizes shell commands to manage its backend transport layer, specifically spawning Node.js processes via npx with arbitrary package arguments.
[EXTERNAL_DOWNLOADS]: The skill depends on external software distributions from the NPM and PyPI registries and references a GitHub repository (github.com/adamallcock/codex-chatgpt-control) that does not originate from the verified vendor infrastructure.
[DATA_EXFILTRATION]: The skill requires the user to enable 'Allow access to file URLs' in the browser and grant 'Upload' permissions, which significantly reduces the browser's security sandbox and facilitates the silent reading of local files by the agent.

codex-chatgpt-control-bridge