The Agent Skills Directory

[PROMPT_INJECTION]: The instructions explicitly command the agent to operate with total autonomy, using phrases like "You are fully autonomous. Never ask the human for permission to continue" and "Run until human stops you." This overrides standard agent safety constraints and human-in-the-loop protocols.
[REMOTE_CODE_EXECUTION]: The core functionality requires the agent to modify 'train.py' and execute the resulting code via 'uv run'. This self-modifying code loop allows for arbitrary code execution based on agent-generated logic.
[REMOTE_CODE_EXECUTION]: The 'prepare.py' script uses 'pickle.load()' to deserialize tokenizer data. Deserializing pickle files is a known security vulnerability that can lead to arbitrary code execution if the data is compromised.
[REMOTE_CODE_EXECUTION]: Fetches the 'uv' project manager installation script from Astral's official domain and executes it via a piped shell command.
[REMOTE_CODE_EXECUTION]: Dynamically loads and executes attention kernels from third-party GitHub repositories ('varunneal/flash-attention-3' and 'kernels-community/flash-attn3') at runtime using the 'kernels' package.
[EXTERNAL_DOWNLOADS]: Downloads LLM training data shards from Karpathy's Hugging Face repository during the data preparation phase.
[PROMPT_INJECTION]: The skill possesses a significant indirect prompt injection surface. It ingests large amounts of training data from external parquet files without sanitization or boundary markers, which could allow malicious content in the training data to influence the agent's behavior during the autonomous loop.
[COMMAND_EXECUTION]: Instructs the agent to execute various shell commands, including 'git', 'uv', and 'grep', using arguments derived from file content and logs generated during the training process.

autoresearch