The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The suggested command uv run python scripts/math_router.py route "<user's math request>" is highly vulnerable to shell injection. Since the user input is interpolated into the shell string, an attacker can use characters like ;, &, or $() to execute unauthorized system commands.
[PROMPT_INJECTION] (HIGH): This skill exhibits a severe Indirect Prompt Injection surface. It ingests untrusted user input from the request and uses it to drive agent actions.
Evidence Chain (Category 8):
Ingestion points: The <user's math request> field in the uv run command.
Boundary markers: Absent; user input is directly concatenated.
Capability inventory: The agent is explicitly commanded to 'execute the returned command' in step 4, which provides an execution vector for malicious strings.
Sanitization: None provided.
[COMMAND_EXECUTION] (HIGH): The instruction 'You execute the returned command' allows for dynamic code execution where the agent runs potentially malicious commands generated by the routing script without a human-in-the-loop or verification step.

math-router