26zl AI and LLM security review
Use this skill for AI applications, agents, RAG systems, model gateways, prompt chains, evals, and LLM governance.
Review workflow
- Inventory the AI system: model/provider, prompts, tools, RAG sources, memory, logs, user roles, secrets, data classes, and downstream actions.
- Threat model trust boundaries:
- user input to prompt
- retrieved content to model
- model output to tools
- tool output to user
- logs/traces to operators
- Test high-risk paths:
- direct and indirect prompt injection
- data exfiltration from RAG or memory
- insecure tool invocation
- overbroad agent permissions
- jailbreaks that change policy or role
- model/provider key leakage
- training/eval data contamination
- Recommend controls:
- least-privilege tool scopes
- allowlisted tool schemas and argument validation
- retrieval filtering and source attribution
- secret redaction before prompts/logs
- output validation before side effects
- human approval for destructive or external actions
- continuous evals and regression prompts
Deliverables
Return findings as:
| Risk | Attack path | Impact | Evidence | Control | Test to keep fixed |
|---|
When the task involves current AI regulation or sector obligations, verify against current official sources before making definitive claims.