The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to interpolate user-provided values directly into JSON payloads for shell commands, creating an attack surface for indirect prompt injection or command injection. 1. Ingestion points: The skill collects user_id, max_budget, user_role, and models from user input as described in SKILL.md. 2. Boundary markers: No delimiters or explicit instructions to treat user data as untrusted are present in the interpolation templates. 3. Capability inventory: The skill is permitted to execute curl via Bash according to the allowed-tools frontmatter. 4. Sanitization: The instructions do not specify any sanitization, escaping, or validation requirements for the agent to apply to the user-supplied strings.
[COMMAND_EXECUTION]: The skill relies on the execution of curl commands in a Bash environment to perform API operations on the proxy server as its primary intended functionality.

update-user