The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (HIGH): The skill implements a system to compile and execute JavaScript and Python code from natural language strings (e.g., guard_js, guard_py). This functionality allows for arbitrary code execution within the host environment.- [COMMAND_EXECUTION] (HIGH): The skill metadata explicitly allows read_file and write_file tools. When combined with dynamic execution, this allows an attacker to read sensitive data or write malicious files to the filesystem.- [PROMPT_INJECTION] (HIGH): The skill converts untrusted natural language into executable code via the COMPILE_EXPRESSION mechanism. Evidence Chain: (1) Ingestion Points: source field in COMPILE_EXPRESSION event; (2) Boundary Markers: Absent; (3) Capability Inventory: File system access (read_file, write_file) and code execution; (4) Sanitization: Absent.- [DATA_EXFILTRATION] (MEDIUM): Utility functions like world.log and world.emit provide a path to leak data retrieved from the filesystem to the agent's output stream, which could then be exfiltrated.

context