// Free tool · 2 minutes · answers stay in your browser
Will your AI feature pass security review?
Ten questions — the same checks a security reviewer runs. Get your score, see exactly what will be flagged, and leave with a punch list. Answers never leave your browser.
Question 1 of 10: Do you have a golden-set eval — curated test inputs with expected behavior — that runs on every prompt or model change?
01 / 10
EVAL
Do you have a golden-set eval — curated test inputs with expected behavior — that runs on every prompt or model change?
Nothing you answer leaves this page — the score is computed in your browser.
// What the scorecard checks
The 10 checks security reviewers actually run
These are the dimensions that decide whether a GenAI feature ships or stalls. They come from real reviews in regulated environments — fintech, identity, payments — not from a framework slide.
01 · EVAL
A repeatable evaluation set
Do you have a golden-set eval — curated test inputs with expected behavior — that runs on every prompt or model change?
Why it matters: Without evals, “how do you know it won’t hallucinate in prod?” has no answer on paper — the first question security asks.
02 · GATE
Evals that gate releases
Would a failing eval actually block a deploy?
Why it matters: An eval that can’t stop a bad deploy is a report, not a control — reviewers distinguish the two.
03 · TRUTH
Groundedness checks on retrieved answers
If your feature answers from data (RAG or retrieval), are outputs checked for grounding against their sources?
Why it matters: Ungrounded answers are confident fabrications — the failure mode that makes legal and security veto launches.
04 · INJECT
Prompt-injection red-teaming
Have you red-teamed prompt injection — including indirect injection through retrieved documents or tool output?
Why it matters: Security flags the LLM as an open door: untrusted content that can become instructions is an exploit path.
05 · PII
PII minimization before prompts
Is personal data redacted or minimized before prompts leave your boundary?
Why it matters: Sending raw PII to a third-party model without minimization is a privacy finding on its own.
06 · DPA
Provider data terms in writing
Do you have your AI provider’s data-retention and training-opt-out terms in writing (DPA / zero-data-retention)?
Why it matters: “Where does our data go?” needs a document, not a vendor’s marketing page.
07 · AUDIT
A reconstructable audit trail
Can you reconstruct any model decision after the fact — prompt, retrieval, and output?
Why it matters: Compliance needs to reconstruct decisions; if you can’t replay an interaction, you can’t investigate one.
08 · SCOPE
Least-privilege tool access
If your AI calls tools or APIs, are those tools least-privilege, allowlisted, and rate-limited?
Why it matters: An agent with broad tool access turns one bad completion into real-world damage.
09 · KILL
A production kill switch
Can you disable the AI path in production instantly — without a deploy?
Why it matters: When the model misbehaves at 2am, “we’ll ship a fix” is not an incident response.
10 · DOC
A written approval evidence pack
Do you have a written document a security reviewer could actually approve from — evals, guardrails, data flows?
Why it matters: Reviewers approve documents, not vibes. No evidence pack means the review stalls by default.