The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains extensive lists of prompt injection, jailbreaking, and token smuggling templates in llm-attacks.md. These are provided as payload examples for users to test against external challenge targets and are not instructions designed to subvert the agent's own behavior or safety filters.
[EXTERNAL_DOWNLOADS]: The skill recommends installing well-known machine learning libraries such as torch, transformers, and scikit-learn from official package registries (PyPI).
[COMMAND_EXECUTION]: Includes Bash and Python one-liners for inspecting local model files and sending HTTP requests to challenge endpoints. These operations are transparent and consistent with the skill's stated purpose of solving AI-related security puzzles.
[DYNAMIC_EXECUTION]: Employs torch.load and safetensors to analyze machine learning models. Users should be aware that loading untrusted .pt files (which use the pickle module) can result in arbitrary code execution, though this is a standard and necessary practice in ML security research.

ctf-ai-ml