The Agent Skills Directory

[PROMPT_INJECTION]: The reference file references/llm-ops-guardrails.md contains regex patterns and strings for detecting prompt injection techniques like 'DAN' and 'ignore instructions'. These are implemented as defensive logic within a guardrail system to protect production LLMs rather than as malicious instructions to the agent itself.
[COMMAND_EXECUTION]: references/fine-tuning.md and references/model-serving.md include shell commands for launching distributed training with torchrun and serving models with vllm. These represent standard industry practices for the documented LLMOps workflows.
[EXTERNAL_DOWNLOADS]: The skill references downloading model weights (e.g., Llama-3.1 models) from Hugging Face and using official Docker images for serving frameworks. These are well-known and trusted sources within the AI development community.
[DATA_EXFILTRATION]: Network operations are utilized in references/model-serving.md to route requests to primary and backup inference providers using the httpx library. This is the intended primary functionality for a model serving gateway.
[CREDENTIALS_UNSAFE]: references/llm-ops-guardrails.md contains a PIIGuardrail class that includes regex patterns for detecting hardcoded secrets such as API keys (sk-). This is implemented as a security scanner to prevent the accidental exposure of credentials in prompts, not as an exposure of the skill author's own secrets.

ai-llm-ops