The Agent Skills Directory

[PROMPT_INJECTION]: The skill metadata in SKILL.md identifies the author as "K-Dense Inc.", which conflicts with the provided author context "DenDen047". This discrepancy in authorship is a deceptive metadata finding.
[PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection by processing untrusted data from reinforcement learning environments. 1. Ingestion points: Environment observations and metadata received through the step() and reset() methods in scripts/env_template.py and various documentation files. 2. Boundary markers: No delimiters or instructions are used to distinguish environment data from agent instructions. 3. Capability inventory: The skill can execute complex shell commands via torchrun and the puffer CLI, and can perform file system writes for checkpoints. 4. Sanitization: No validation or sanitization of environment data is performed before it is processed or logged.
[COMMAND_EXECUTION]: The skill relies on powerful command-line utilities such as torchrun and the puffer CLI for managing distributed multi-process training tasks across local or remote nodes.

pufferlib