Who this is for
Engineering and product teams building AI agents, RAG systems, or any feature that puts an LLM in front of untrusted input or gives it access to tools and data. If your product calls a model API and does anything with the response beyond displaying it as plain text, this toolkit applies.
What's in the toolkit
- Model access control — who and what can call the model, and with what permissions.
- Data exposure — what context the model sees, and what it's allowed to leak.
- Prompt injection surface — direct injection via user input, and indirect injection via retrieved documents or tool outputs.
- Output validation — treating model output as untrusted input to the next step, not as a trusted instruction.
- Governance — vendor API key hygiene, logging, and who owns the failure mode when the model is wrong.
How to use it
Start with the AI Security Checklist for the product-level view, then the Prompt Injection Checklist for the specific attack surface — most real incidents trace back to a gap this second checklist would have caught.
Go deeper
For a broader product-security engagement, see AI Security Kenya. For RAG-pipeline-specific and autonomous-agent-specific security, see RAG Security and AI Agent Security. For the technical deep-dive on injection mechanics, see Prompt Injection and LLM Security.