docs: describe relay metadata and audit status precisely

State what the relay sees at runtime (authenticated submission reveals
which device hands an envelope to which address) instead of claiming it
never learns the sender, note where data is stored, drop the
"Signal-grade" label and add a non-affiliation notice, and document
that only internal audits and penetration tests have been done so far.

SECURITY.md

@@ -28,7 +28,16 @@ reproduce almost everything.
 
 ## What the relay can and cannot see
 
-The threat model is documented in the README: the relay sees routing tokens,
-message sizes, timing and push tokens, never content or sender identity.
-Reports that demonstrate the relay learning more than that are the ones we
-care about most.
+The threat model is documented in the README. In short: envelopes carry no
+sender (sealed sender) and the relay never sees content. It does see routing
+tokens, envelope sizes, timing and push tokens, and while a connection is
+open the client IP and which authenticated device submits an envelope to
+which delivery address. It stores no sender–recipient link and no IP
+addresses. Reports that demonstrate the relay learning or keeping more than
+that are the ones we care about most.
+
+## Audits
+
+We run regular internal security audits and penetration tests (August,
+September and October 2026); fixes are listed in `CHANGELOG.md`. No
+independent third-party audit has taken place yet.