The Agent Skills Directory

[METADATA_POISONING]: The skill's description and body claim that 'torchforge is Meta's PyTorch-native RL library.' However, the referenced resources point to 'meta-pytorch.org' and 'github.com/meta-pytorch,' which are not recognized as official domains or repositories for Meta or the PyTorch project. This indicates potential impersonation designed to mislead users regarding the software's origin and safety.
[EXTERNAL_DOWNLOADS]: The instructions encourage users to download code from unverified external sources (e.g., github.com/meta-pytorch/torchforge) and execute local installation scripts (./scripts/install.sh) that rely on these untrusted repositories.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to perform high-privilege operations, such as executing installation scripts, running distributed training processes, and submitting Slurm jobs (sbatch). These actions are performed using a library from an unverified source that is masquerading as a major vendor.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external datasets from Hugging Face (openai/gsm8k) within its training workflows. The instructions lack boundary markers or sanitization requirements for this untrusted data, creating a surface where malicious content in the training data could attempt to influence the agent's behavior during high-privilege execution phases.
Ingestion points: Dataset defined in SKILL.md (e.g., path: 'openai/gsm8k').
Boundary markers: Absent; no instructions for using delimiters or ignoring embedded commands.
Capability inventory: Shell script execution, multi-process Python training, and Slurm job submission (SKILL.md).
Sanitization: Absent; no instructions for validating or escaping dataset content.

torchforge-rl-training