The Agent Skills Directory

Indirect Prompt Injection (LOW): Multiple modules describe patterns where untrusted external data is interpolated into LLM prompts without robust sanitization.
Ingestion points: Data enters the context via user feedback strings in user-feedback/SKILL.md, retrieved documents in rag-systems/SKILL.md, and tool execution results in ai-agents/SKILL.md.
Boundary markers: While the ai-agents/SKILL.md module uses an Observation: marker in the ReAct loop, most other modules lack explicit delimiters (e.g., XML tags or JSON schemas) to isolate untrusted content from instructions.
Capability inventory: The patterns described include high-privilege capabilities such as tool execution in ai-agents/SKILL.md, and automated finetuning/deployment logic in user-feedback/SKILL.md.
Sanitization: A regex-based sanitize_input function is demonstrated in prompt-engineering/SKILL.md for defense, but the skill does not consistently apply this or other sanitization methods to the data ingestion examples in other modules.
Prompt Injection (SAFE): The prompt-engineering/SKILL.md and guardrails-safety/SKILL.md files contain explicit prompt injection strings (e.g., 'ignore previous instructions'). These are used exclusively for educational purposes to demonstrate detection and defensive prompting techniques.

ai-engineering