The Agent Skills Directory

[COMMAND_EXECUTION]: The skill recommends using common system utilities like file, xxd, head, and awk for file inspection. These are standard diagnostic tools used here for their intended purpose of format analysis.\n- [PROMPT_INJECTION]: Indirect prompt injection surface detected: (1) Ingestion point: Skill instructs on reading and parsing external data files (SKILL.md). (2) Boundary markers: Absent. (3) Capability inventory: Recommends shell command execution and ML model loading on data. (4) Sanitization: No sanitization methods for file content are described.\n- [REMOTE_CODE_EXECUTION]: Mention of torch.load() for processing model checkpoints introduces a risk of unsafe deserialization via the pickle module. However, the skill provides this as documentation of common practices and also mentions safer alternatives like safetensors.

data-formats