AI Security Testing Cheat Sheet
20 sections · 176 techniques · Free
AI Security Testing Cheat Sheet
A practical methodology for testing LLM apps, RAG pipelines, AI agents, and MCP tools — grouped by where issues surface, with controlled payloads and a result tracker you can export as findings.
- OWASP LLM Top 10 · 2026
- RAG pipelines
- AI agents
- MCP tools
Start here
Attack surface mapping
Make the AI trust boundaries visible before testing.
Prompt injection
The core input attack — data vs. instructions.
Payload encoder
Obfuscate payloads to slip past keyword filters.
Reporting & severity
Translate AI behavior into AppSec impact.
How it works
Every section targets a place where untrusted content or model output can cross a trust boundary. Work the checklist by signal — the observable proof a technique worked — not by whether a prompt sounds clever.
Three anchors for every test
Trust boundary
Keep user and retrieved content as data.
Control point
Enforce permissions outside the model.
Evidence
Record impact, reproduction, and remediation.
The signal map
Where issues surface, input to output. Start anywhere — each stage links to its sections.
Stage 01 Inputs chat, files, tickets, web Stage 02 Retrieval RAG, indexes, memory Stage 03 Model instructions, context, policy Stage 04 Tools APIs, agents, MCP Stage 05 Outputs UI, email, code, exports
Risk by capability
Risk scales with capability
A public text-only chatbot and an agent with write, email, or admin tools are not the same test — scope effort to what the AI can actually reach or do. See the full risk-by-capability table on Attack surface mapping →
Track your review
1
Mark
Flag each technique Pass, Vuln, or N/A as you test.
2
Save
Results and notes persist in your browser — no account needed.
3
Export
Download a Markdown report of everything flagged vulnerable, grouped
by section with a suggested severity.