The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits two integrity concerns. First, eight large reference files (including architecture-patterns.md, security-best-practices.md, and state-management.md) consist of approximately 150 repetitive sections of generic technical filler text. This massive volume of repetitive data acts as a form of noise injection that obscures potential malicious payloads and can confuse the agent. Second, the skill is vulnerable to indirect prompt injection. Ingestion points: Training parameters and metrics logged from external training processes in SKILL.md and experiment-tracking-fundamentals.md. Boundary markers: Absent. Capability inventory: Uses subprocess.check_output to execute git and dvc commands in SKILL.md and experiment-reproducibility.md. Sanitization: None provided for external inputs logged as parameters or metrics.\n- [COMMAND_EXECUTION]: The skill instructions and reference files (experiment-reproducibility.md) include functions that use subprocess.check_output to run shell commands like 'git rev-parse HEAD' and 'dvc status'. While used for legitimate reproducibility purposes, executing arbitrary shell commands based on strings derived from the local environment or project data poses a security risk if not strictly constrained.

ml-experiment-tracking