diff --git a/plugins/magpie-security/hooks/public-surface-scrub.test.ts b/plugins/magpie-security/hooks/public-surface-scrub.test.ts
index d01db8433..4bde063fc 100644
--- a/plugins/magpie-security/hooks/public-surface-scrub.test.ts
+++ b/plugins/magpie-security/hooks/public-surface-scrub.test.ts
@@ -2443,4 +2443,52 @@ upstream_repo: apache/kafka
     });
   });
 
+  describe('a blocked gh api write names its vetted route', () => {
+    const projectConfig = { trackerRepo: 'airflow-s/airflow-s', upstreamRepo: 'apache/airflow' };
+    const evaluate = (cmd: string) =>
+      evaluatePublicSurfaceCommand(cmd, { projectConfig, readFile: async () => 'Fixed as CVE-2026-1234.' });
+
+    test('an advisory comment post is still blocked and points at advisory-comment', async () => {
+      const res = await evaluate(
+        'gh api -X POST repos/apache/airflow/security-advisories/GHSA-abcd-efgh-ijkl/comments -F body=@/tmp/b.md',
+      );
+      assertStrictEqual(res.blocked, true);
+      assert(res.reason?.includes('advisory-comment GHSA-abcd-efgh-ijkl <body-file>'), res.reason);
+      assert(res.reason?.includes('[callers]'), res.reason);
+    });
+
+    test('an advisory field edit is still blocked and says no vetted operation exists', async () => {
+      const res = await evaluate('gh api --method PATCH repos/apache/airflow/security-advisories/GHSA-abcd-efgh-ijkl -f cve_id=CVE-2026-1234');
+      assertStrictEqual(res.blocked, true);
+      assert(res.reason?.includes('No vetted operation edits advisory record fields'), res.reason);
+    });
+
+    test('an advisory comment edit says to use the web UI', async () => {
+      const res = await evaluate('gh api -X PATCH repos/apache/airflow/security-advisories/GHSA-abcd-efgh-ijkl/comments/12 -F body=@/tmp/b.md');
+      assertStrictEqual(res.blocked, true);
+      assert(res.reason?.includes('No vetted operation edits or deletes an advisory comment'), res.reason);
+    });
+
+    test('a comment edit and a comment post point at comment-update and issue-comment', async () => {
+      const edit = await evaluate('gh api -X PATCH repos/apache/airflow/issues/comments/987 -F body=@/tmp/b.md');
+      assertStrictEqual(edit.blocked, true);
+      assert(edit.reason?.includes('comment-update 987 <body-file>'), edit.reason);
+
+      const post = await evaluate('gh api repos/apache/airflow/issues/42/comments -F body=@/tmp/b.md');
+      assertStrictEqual(post.blocked, true);
+      assert(post.reason?.includes('issue-comment 42 <body-file>'), post.reason);
+    });
+
+    test('other gh api writes get no vetted-route hint', async () => {
+      const res = await evaluate('gh api -X POST repos/apache/airflow/pulls/5/comments -f body="Fixes CVE-2026-1234"');
+      assertStrictEqual(res.blocked, true);
+      assert(!res.reason?.includes('vetted-op'), res.reason);
+    });
+
+    test('the tracker exemption is unchanged', async () => {
+      const res = await evaluate('gh api -X PATCH repos/airflow-s/airflow-s/issues/comments/987 -F body=@/tmp/b.md');
+      assertStrictEqual(res.blocked, false, JSON.stringify(res));
+    });
+  });
+
 });
diff --git a/plugins/magpie-security/hooks/public-surface-scrub.ts b/plugins/magpie-security/hooks/public-surface-scrub.ts
index 387574f54..02cf05d92 100644
--- a/plugins/magpie-security/hooks/public-surface-scrub.ts
+++ b/plugins/magpie-security/hooks/public-surface-scrub.ts
@@ -677,6 +677,10 @@ export interface ParsedSubcommand {
   hasDynamicInput: boolean;
   isGraphqlMutation?: boolean;
   isLiteralUnknownEndpoint?: boolean;
+  /** For `gh api`: the REST endpoint path, with any host, `v3` prefix and query string stripped. */
+  apiEndpoint?: string;
+  /** For `gh api`: the HTTP method given with `-X` / `--method`, upper-cased. */
+  apiMethod?: string;
   hasGitDirOrWorkTree?: boolean;
   /** The command runs with environment set for it: a leading `NAME=value` or an `env` wrapper, after any prefix. */
   setsOwnEnvironment?: boolean;
@@ -1749,6 +1753,7 @@ export function parseSubcommand(subcommand: string): ParsedSubcommand {
         let endpoint = token.replace(/^(?:https?:\/\/[^/]+)?\/+/, '');
         // Strip optional v3 or api/v3 prefix
         endpoint = endpoint.replace(/^(?:api\/)?v3\/+/, '');
+        result.apiEndpoint = endpoint.replace(/[?#].*$/, '');
 
         if (endpoint.startsWith('repos/')) {
           const parts = endpoint.slice('repos/'.length).split('/');
@@ -1794,6 +1799,17 @@ export function parseSubcommand(subcommand: string): ParsedSubcommand {
         break;
       }
 
+      for (let i = cmdIdx + 2; i < tokens.length; i++) {
+        const token = tokens[i];
+        if ((token === '-X' || token === '--method') && i + 1 < tokens.length) {
+          result.apiMethod = tokens[i + 1].toUpperCase();
+        } else if (token.startsWith('--method=')) {
+          result.apiMethod = token.slice('--method='.length).toUpperCase();
+        } else if (/^-X[A-Za-z]+$/.test(token)) {
+          result.apiMethod = token.slice(2).toUpperCase();
+        }
+      }
+
       for (let i = cmdIdx + 2; i < tokens.length; i++) {
         const token = tokens[i];
 
@@ -2056,6 +2072,66 @@ export function generateBlockingReason(commandType: string, terms: string[]): st
   ].join('\n');
 }
 
+/**
+ * For a blocked `gh api` write that has a vetted route, names that route.
+ *
+ * Advisory discussions and tracker comments are private surfaces whose text legitimately carries
+ * the terms this guard scans for. The guard's decision does not change; the hint only names the
+ * vetted-ops operation that performs the same write, and says the operation has to be granted to
+ * the caller in the adopter's vetted-ops config, which is a maintainer's change, not the agent's.
+ * Returns undefined when the endpoint has no vetted route.
+ */
+export function vettedRouteHint(parsed: ParsedSubcommand): string | undefined {
+  if (parsed.commandType !== 'gh api' || !parsed.apiEndpoint) {
+    return undefined;
+  }
+  const endpoint = parsed.apiEndpoint;
+  const sendsFields = parsed.extractedTexts.length > 0 || parsed.referencedFiles.length > 0;
+  const method = parsed.apiMethod ?? (sendsFields ? 'POST' : 'GET');
+  const grant =
+    'The calling skill needs that operation in its `[callers]` list in ' +
+    '`.apache-magpie-overrides/tools/vetted-ops/config.toml`, and adding it is a maintainer\'s change. ' +
+    'The body file goes under `~/.cache/apache-magpie/vetted-ops-bodies/` (see tools/vetted-ops/README.md).';
+
+  const advisoryComments = endpoint.match(
+    /^repos\/[^/]+\/[^/]+\/security-advisories\/(GHSA-[0-9a-z-]+)\/comments(\/\d+)?\/?$/i,
+  );
+  if (advisoryComments) {
+    if (advisoryComments[2] || method === 'PATCH' || method === 'DELETE') {
+      return 'No vetted operation edits or deletes an advisory comment; do that in the advisory web UI.';
+    }
+    return [
+      'A security advisory discussion is private to its collaborators. Post the comment with the vetted operation instead:',
+      `  vetted-op --caller <skill> advisory-comment ${advisoryComments[1]} <body-file>`,
+      grant,
+    ].join('\n');
+  }
+
+  if (/^repos\/[^/]+\/[^/]+\/security-advisories\/GHSA-[0-9a-z-]+\/?$/i.test(endpoint)) {
+    return 'No vetted operation edits advisory record fields (`cve_id`, `credits`, `severity`, ...); edit them in the advisory web UI.';
+  }
+
+  const commentEdit = endpoint.match(/^repos\/[^/]+\/[^/]+\/issues\/comments\/(\d+)\/?$/);
+  if (commentEdit && method === 'PATCH') {
+    return [
+      'If this comment is on the private tracker, edit it with the vetted operation instead:',
+      `  vetted-op --caller <skill> comment-update ${commentEdit[1]} <body-file>`,
+      grant,
+    ].join('\n');
+  }
+
+  const issueComment = endpoint.match(/^repos\/[^/]+\/[^/]+\/issues\/(\d+)\/comments\/?$/);
+  if (issueComment && method === 'POST') {
+    return [
+      'If this issue is on the private tracker, post the comment with the vetted operation instead:',
+      `  vetted-op --caller <skill> issue-comment ${issueComment[1]} <body-file>`,
+      grant,
+    ].join('\n');
+  }
+
+  return undefined;
+}
+
 /**
  * Commands known to be purely read-only or harmless environment manipulations.
  * If any preceding command on a composite shell line is NOT in this set,
@@ -2499,6 +2575,10 @@ export async function evaluatePublicSurfaceCommand(
     if (matched.length > 0) {
       const uniqueTerms = Array.from(new Set(matched));
       let reason = generateBlockingReason(parsed.commandType, uniqueTerms);
+      const vettedHint = vettedRouteHint(parsed);
+      if (vettedHint) {
+        reason += `\n\n${vettedHint}`;
+      }
       if (trackerFallbackHint) {
         // The Bash tool call carries no cwd, so a command run in the tracker checkout cannot be
         // told apart from one run after an earlier `cd ../<upstream>`; an explicit target can.
diff --git a/plugins/magpie-security/skills/issue-sync/github-advisory.md b/plugins/magpie-security/skills/issue-sync/github-advisory.md
index 66a119085..bb60e04ab 100644
--- a/plugins/magpie-security/skills/issue-sync/github-advisory.md
+++ b/plugins/magpie-security/skills/issue-sync/github-advisory.md
@@ -115,6 +115,8 @@ It returns:
 
 The patch is one Step 4 proposal item per advisory; apply it with
 `gh api --method PATCH repos/<upstream>/security-advisories/<GHSA> --input <patch-file>`.
+Where the public-surface guard runs, it blocks that `PATCH` whenever the patch names a CVE (a `cve_id` always does), and no vetted op edits advisory fields:
+hand the operator the advisory `html_url` and the field values to set in the web UI, and record the change only after re-reading the advisory.
 
 ### Two-way comment sync — `<sm> ghsa-mirror`
 
@@ -137,6 +139,8 @@ Both are proposals: the tracker copies are one Step 4 item, and **each** advisor
 since it reaches the reporter at once and cannot be deleted through the API.
 Apply the copies with `gh issue comment <N> --repo <tracker> --body-file <file>` or the `comment-update` vetted op,
 and the advisory posts with the `advisory-comment` vetted op (or `gh api --method POST …/comments -F body=@<file>`; an edit is `--method PATCH …/comments/<id>`).
+The public-surface guard blocks the plain `gh api` forms when the text names a CVE or the advisory, so where it runs the vetted op is the only route;
+it needs `advisory-comment` in the caller's `[callers]` list of the adopter's vetted-ops config, and when that grant is missing the operator adds it — the agent does not.
 
 ### Read the advisory discussion