The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's core workflow involves 'Extracting the function code' from an LLM response and 'writing to a temp file' to be executed. This pattern facilitates the execution of code generated at runtime, which is a significant security risk if the generated code contains malicious logic.\n- [COMMAND_EXECUTION]: Under the 'How it works' section, the agent is directed to 'Run the oracle (the user-provided test or property check)' against the locally generated file. This instructions the agent to perform shell execution of code that has not been reviewed by a human.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted user specifications and uses them to generate and execute code.\n
Ingestion points: User-provided task specifications (function signatures and algorithm descriptions) mentioned in the 'When to invoke' section of SKILL.md.\n
Boundary markers: The skill mentions using a 'strict prompt template' to control the local model's output, but it lacks clear delimiters to isolate untrusted user input within that template.\n
Capability inventory: The skill explicitly authorizes the agent to write files ('write to a temp file') and execute shell commands ('Run the oracle') as documented in the 'How it works' section of SKILL.md.\n
Sanitization: There is no mention of sanitizing, sandboxing, or validating the model-generated code before it is executed on the local system.

concord-microtask-router