The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The primary workflow of this skill involves the agent executing shell commands that are dynamically generated by the math_router.py script. The documentation explicitly instructs the agent to 'execute the returned command,' which facilitates a direct path from script output to shell execution without manual verification.
REMOTE_CODE_EXECUTION (MEDIUM): The skill enables the execution of dynamically constructed commands triggered by untrusted user input. If the math_router.py script or the agent's shell wrapper does not properly sanitize the <user's math request>, an attacker could inject shell metacharacters to execute arbitrary code on the host system.
PROMPT_INJECTION (LOW): The skill instructions include high-priority directives ('ALWAYS use this router first') intended to override the agent's default decision-making process for tool selection, which could be used to force the agent into a vulnerable execution path.
INDIRECT_PROMPT_INJECTION (LOW):
Ingestion points: User-provided math requests are ingested directly as command-line arguments to the math_router.py script.
Boundary markers: The instructions suggest using double quotes ("<user's math request>") but do not provide guidance on escaping or sanitization of those inputs.
Capability inventory: The agent has the capability to run shell commands via uv run python and is directed to execute the results of the routing script.
Sanitization: No evidence of input validation or sanitization is present in the skill instructions.

math-router