The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Automated scans detected piped execution to Python via curl -s http://localhost:$PORT/v1/models | python3 in scripts/deploy.sh and SKILL.md. These are localized health checks and model ID extraction tasks performed against the user's own inference server running on localhost. They do not involve external, untrusted sources.
[EXTERNAL_DOWNLOADS]: The skill downloads checkpoints and tokenizer configurations from HuggingFace Hub (e.g., nvidia/Llama-3.1-8B-Instruct-FP8). These references target trusted NVIDIA-managed repositories and well-known services.
[COMMAND_EXECUTION]: The skill makes extensive use of shell commands and nohup to manage long-running inference server processes. It also uses sed to patch framework source code in references/unsupported-models.md for compatibility fixes, which is a documented iterative debug practice for this specific domain.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes untrusted user prompts and external model checkpoints. However, it implements standard boundary markers for testing (e.g., specific coherence gate checks in references/benchmarking.md) and the risk is mitigated by the intended purpose of the skill (inference serving).

deployment