[kmcginnes]

Summary

Security reviews keep reporting behavior that's by design, such as no built-in auth or the proxy signing with the host's shared AWS identity, because REVIEW.md never pointed reviewers at the threat model in docs/references/security.md.

User-facing: none

 ## What Important means here
+  - Security changes that widen a documented risk or break a guarantee
+## Security model
+  - docs/references/security.md is the threat model; test and judge against it
+  - deployer-owned behavior is by design
+  - "never / does not / refuses" statements are guarantees; breaking one is Important
+  - Graph Explorer owns untrusted input
 ## Verification bar
+  - a broken guarantee names the doc statement instead of a triggering input
 ## Do not report
-  - Security behavior that's by design: (3 bullets)
+  - Security behavior that's by design, per the Security model section

The section points at the doc instead of copying its by-design list, so the two can't drift.

Merge Danger

Door: two-way

Impact: reviews

Changes how review agents and the PR review bot judge security findings. No code or docs outside REVIEW.md change.

Related

• Related to Document the supported hosting model #2459

---

_{🤖 Written by an AI agent on my behalf.}

[kmcginnes]

Approved