The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands provided in the <run_cmd> binding. It instructs the agent to run these commands in a loop without human oversight.
[REMOTE_CODE_EXECUTION]: The skill's primary purpose is to dynamically 'hack' (modify) Python training scripts and then execute them. This constitutes a controlled form of dynamic code generation and execution. While limited to <editable_files>, an agent could potentially inject malicious logic into these files which is then executed by the <run_cmd>.
[PROMPT_INJECTION]: The skill contains strong behavioral overrides such as 'LOOP FOREVER', 'NEVER STOP', and instructions to 'NOT pause to ask the human if you should continue,' which explicitly attempts to bypass the agent's typical interaction safety pattern of seeking confirmation for repetitive or impactful actions.

karpathy