Pre-launch safety checklist
Mapped to the OWASP Top 10 for LLM Applications (2025).
QAProductDeveloperDesigner
OWASP LLM 2025 map
| ID | Risk | Check |
|---|---|---|
| LLM01 | Prompt injection | Untrusted content as data; least privilege; confirmations |
| LLM02 | Sensitive information disclosure | Redaction, retention limits, access control |
| LLM03 | Supply chain | Vet models, servers and dependencies |
| LLM04 | Data and model poisoning | Control training and ingestion sources |
| LLM05 | Improper output handling | Validate before downstream use |
| LLM06 | Excessive agency | Narrow tools; approvals; separate environments |
| LLM07 | System prompt leakage | No secrets in prompts |
| LLM08 | Vector and embedding weaknesses | Access filters; ingestion review |
| LLM09 | Misinformation | Grounding, citations, not-found path |
| LLM10 | Unbounded consumption | Rate, token and cost caps |
Launch gate
- Disclosed as AI; human option offered
- Not-found path and handoff tested
- Adversarial suite run; no critical findings open
- Kill switch rehearsed
- Logs redacted; owner and incident process named
- Rules for your markets checked with qualified advisors