The Agent Skills Directory

[DYNAMIC_EXECUTION]: The classifier.py module uses pickle.load() to deserialize a local model file. This is an unsafe practice as it allows for arbitrary code execution if the file is tampered with.
[PERSISTENCE_MECHANISMS]: The skill provides an install-hook command that modifies the .git/hooks directory to install a pre-commit hook. This constitutes a persistence mechanism that executes code during git operations.
[COMMAND_EXECUTION]: The skill frequently executes external processes and other skills via subprocess.run, including /memory, /assistant, and system utilities like git and memory-agent. This increases the overall attack surface.
[PROMPT_INJECTION]: The skill ingests untrusted conversation logs for analysis by machine learning models and LLMs. There is a lack of boundary markers or sanitization in scripts like llm_auditor.py, making it vulnerable to indirect prompt injection. Ingestion points: detect and report commands in lie_detector.py. Boundary markers: Absent. Capability inventory: Numerous subprocess.run calls across the logic. Sanitization: Absent.

lie-detector