The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides instructions to grant containers the SYS_ADMIN capability and use sudo to remove system packages which provides high-level control over the host environment. Evidence in SKILL.md and references/multi-node-training.md.
[INDIRECT_PROMPT_INJECTION]: Documentation in references/custom-rewards.md outlines a process for reward functions that take untrusted model outputs and execute them via subprocesses without any depicted security boundaries or sanitization. Ingestion: queries argument. Capability: subprocess.run.
[DYNAMIC_EXECUTION]: Examples in the skill show how to write model-generated strings to files and execute them as Python scripts at runtime. Evidence: reward_func_code_gen.py logic in references/custom-rewards.md.

openrlhf-training