diff --git a/scripts/check-advisories.py b/scripts/check-advisories.py
index 0800e2e47..f47fb0987 100755
--- a/scripts/check-advisories.py
+++ b/scripts/check-advisories.py
@@ -71,22 +71,49 @@ ALLOW = {
     "RUSTSEC-2026-0368": "zip: extract-only; nothing here extracts an archive",
     "RUSTSEC-2026-0369": "zip: extract-only; nothing here extracts an archive",
     "RUSTSEC-2025-0168": "zip: extract-only path canonicalization; nothing here extracts",
-    # The Marvin attack recovers an RSA key from the timing of private-key
-    # decryption. There is no fixed release of `rsa` at all. It reaches
-    # sandbox-rs and wai-rs transitively and neither holds an RSA private key
-    # nor decrypts: openpay's unused direct dependency was removed outright.
-    "RUSTSEC-2023-0071": "rsa: no fixed release; no private-key decryption here",
+    # The Marvin attack recovers an RSA key from the timing of **private-key
+    # decryption** (PKCS#1 v1.5). There is no fixed release of `rsa` at all, so
+    # this one cannot be bumped away and the reason has to carry it. Both
+    # remaining copies were checked by call site, not by assumption:
+    #
+    #   sandbox-rs: `sandbox-tee/src/amd_kds.rs` is the only direct user in the
+    #     tree. It verifies AMD's KDS chain (ARK -> ASK -> VCEK, RSASSA-PSS /
+    #     SHA-384) with `RsaPublicKey` + `pss::VerifyingKey` + `Verifier`.
+    #     Public-key verification only; there is no decryption anywhere, and
+    #     `RsaPrivateKey` appears solely inside `#[cfg(test)] mod tests`, which
+    #     mints throwaway keys to build a synthetic chain.
+    #   wai-rs: reaches it only through `c2pa`, which is an optional dependency
+    #     behind the `c2pa_emit` feature and off by default, so a default build
+    #     does not compile it. WAI's own signing is Ed25519 (`ed25519_dalek`),
+    #     which uses neither `rsa` nor `crypto-bigint`.
+    #
+    # openpay's copy is gone: the dependency was declared and never used, so it
+    # was removed outright rather than excused.
+    "RUSTSEC-2023-0071": "rsa: no fixed release; verification only, no decryption, private keys test-only",
     # ── Pinned by a dependency; needs an upstream release ──────────────────
     # Secret-dependent branches in operations meant to be constant-time
     # (subtraction, zero counting, division, Montgomery multiplication), so
-    # timing can leak intermediate arithmetic on secret integers. Fixed in
-    # >=0.7.6, and smart-byte's 0.7.x copy is patched. The 0.5.5 line is pinned
-    # by elliptic-curve 0.13, so reaching it means moving the whole
-    # p256/elliptic-curve stack — a real upgrade, not a bump, and this entry is
-    # that debt, recorded on paths that sign and derive keys. Version-pinned on
-    # purpose: a 0.7.x copy must never be excused by it.
+    # timing can leak intermediate arithmetic on a secret integer. Fixed in
+    # >=0.7.6, and smart-byte's 0.7.x copy is patched.
+    #
+    # Every remaining 0.5.5 copy arrives under `elliptic-curve` — it is the only
+    # dependent in all five lockfiles — which pins the 0.5 line, so reaching
+    # >=0.7.6 means moving the whole p256/elliptic-curve stack. That is an
+    # upgrade, not a bump, and this entry is that debt.
+    #
+    # What bounds it: a timing attack of this class recovers a key by observing
+    # many operations under the *same* secret, and none of our own signing runs
+    # through this crate. WAI, sandbox and jouleclaw sign with Ed25519
+    # (`ed25519_dalek` / curve25519-dalek, which does not use crypto-bigint);
+    # in wai-rs the elliptic-curve path arrives under `c2pa`, optional and off
+    # by default, where production callers supply their own signer and the
+    # in-tree `EphemeralSigner` holds a per-call key that reaches
+    # `ValidationState::Valid` and never `Trusted`. So the debt is real but it
+    # is not sitting under a long-lived key of ours.
+    #
+    # Version-pinned on purpose: a 0.7.x copy must never be excused by it.
     "RUSTSEC-2026-0387": {
-        "why": "crypto-bigint 0.5.5 is pinned by elliptic-curve 0.13; 0.7.x is patched",
+        "why": "crypto-bigint 0.5.5 pinned by elliptic-curve 0.13; our signing is Ed25519; 0.7.x is patched",
         "only": ("crypto-bigint 0.5.5",),
     },
     # hpke-rs-libcrux pins `=0.7.0`, which pins libcrux-kem 0.0.9. Both are