The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes rules that instruct the agent to prompt the user to "re-run with network access or auto-approve" if environment restrictions like sandboxing block tools like curl or bash. This is an attempt to bypass security constraints imposed by the execution environment.- [REMOTE_CODE_EXECUTION]: The skill dynamically fetches "Procedure+Rules" from an external vendor-controlled catalog (https://basic-anteater-667.convex.site/capabilities.json and subsequent .md files) and treats the content as instructions to be followed by the agent at runtime.- [EXTERNAL_DOWNLOADS]: Fetches JSON and Markdown configuration data from a vendor-specific deployment domain to determine functionality and procedures.- [INDIRECT_PROMPT_INJECTION]:
Ingestion points: Dynamically fetches markdown instruction files from https://basic-anteater-667.convex.site/capability/<id>.md based on user requests.
Boundary markers: None; the agent is explicitly instructed to follow the procedure and rules found in the external document.
Capability inventory: The skill has access to internal scripts (/add-hosting, /add-component) and shell commands (curl, bash).
Sanitization: None; the skill relies on the agent's "normal judgment" to evaluate fetched instructions instead of technical validation or filtering.

convex-add