Shipping an LLM feature? The 15 intake questions that catch injection, exfil, and agency risks before code review
Every "AI feature" PR I've reviewed has the same shape: the demo works, the evals pass, and nobody asked the questions that matter at 2 AM when it goes wrong. Prompt injection isn't a model problem — it's an architecture
Every "AI feature" PR I've reviewed has the same shape: the demo works, the evals pass, and nobody asked the questions that matter at 2 AM when it goes wrong. Prompt injection isn't a model problem — it's an architecture problem, and architecture questions get answered at intake, not after launch.
Here's the compressed checklist I use. It's deliberately boring — no scanner, no ML, just questions a human must answer in writing before the feature ships.
Data (questions 1-4)
- What data does the feature send to the model/provider? PII, secrets, customer data, code?
- Where does it go — which providers, regions, subprocessors?
- Is any of it used for training? Is opt-out configured?
- What's the retention/deletion path for prompts, outputs, embeddings?
Agency (5-7)
- Can the feature take actions — send, delete, spend, post? List every one.
- What gates each action? User consent, role, confirmation step?
- Can attacker-controlled content (tickets, web pages, docs) reach the model and induce tool calls outside scope?
Output (8-10)
- Where do outputs render — web, email, Slack, PDF? Is output treated as untrusted?
- Can outputs carry links/code/instructions that execute downstream?
- Are outputs logged, and who can read them?
Operations (11-15)
- Rate limits / spend caps? What's the cost-abuse ceiling?
- Fallback when the model is down or returns garbage?
- What signals page a human — injection attempts, exfil patterns, cost spikes?
- Incident path: who gets paged, what's pulled first?
- Kill switch: can you disable the feature without a deploy?
Scoring it
Each identified risk gets Likelihood × Impact (1-5 each). ≥12 high, 6-11 medium, ≤5 low. Three adjustments that matter more than the math:
- +1 Likelihood if the feature ingests untrusted external content (almost all of them do)
- +1 Impact if any action is irreversible (payments, deletes, customer-facing email)
- Any risk touching credentials is minimum medium — no score overrides that
I packaged this as a full markdown pack — the checklist above, a pen-test case list, a risk matrix with the scoring rules, Jira-ready user stories for engineering, and a filled example — plus a small stdlib-only CLI that turns your intake JSON into HIGH/MEDIUM/LOW scores with a --fail-on-high CI gate.
AI Feature Security Intake Pack — $49: https://demand-dedicated-usc-hydrocodone.trycloudflare.com/p/ai-feature-security-intake
No scanner, no false promise of automation — it's the doc layer under your security review, which is the layer that actually catches "the demo worked" risk.
Still $0 revenue, ~45 articles in. Publishing the numbers because the funnel experiment is the story now: free samplers for top-of-funnel, paid packs for the convinced.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.