The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs users to download assets from meta-pytorch.org and the meta-pytorch GitHub organization. These sources are not official and utilize brand-jacking techniques to mimic trusted entities like Meta and the PyTorch Foundation.
[REMOTE_CODE_EXECUTION]: Users are instructed to clone the monarch library from an unverified GitHub repository (github.com/meta-pytorch/monarch) and install it, allowing for the execution of arbitrary code from an untrusted source.
[COMMAND_EXECUTION]: The installation process involves running shell scripts (./scripts/install.sh and ./scripts/install_rocm.sh) found within the skill's file structure or referenced repositories. These scripts are not audited and can perform unauthorized actions on the host system.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its handling of external data.
Ingestion points: Training datasets are loaded from external sources like openai/gsm8k on HuggingFace, and model-generated responses are processed for reward calculation.
Boundary markers: There are no delimiters or specific instructions provided to the agent to treat ingested data as untrusted or to ignore embedded commands.
Capability inventory: The skill has the ability to execute shell commands, manage multi-process training jobs, and write checkpoints to disk.
Sanitization: No validation, escaping, or filtering is applied to the data retrieved from external sources before it influences the training logic or agent behavior.

torchforge-rl-training