diff --git a/actions/setup/js/sanitize_content.test.cjs b/actions/setup/js/sanitize_content.test.cjs
index f139f840f69..5e3bc4d7636 100644
--- a/actions/setup/js/sanitize_content.test.cjs
+++ b/actions/setup/js/sanitize_content.test.cjs
@@ -2425,6 +2425,89 @@ describe("sanitize_content.cjs", () => {
       // Idempotency: a second pass on the decoded result should not re-introduce >
       expect(twice).toBe(once);
     });
+
+    describe("named invisible-character entities — @mention bypass prevention", () => {
+      // These tests cover the bypass described in gh-aw#24154 / gh-aw-security#2086.
+      // Named entity forms of invisible characters must be decoded before Step 3
+      // strips the resulting code points, otherwise the "&" character prevents
+      // neutralizeAllMentions from matching the @ sign.
+
+      it("should decode ­ (soft hyphen U+00AD) and neutralize @mention", () => {
+        expect(sanitizeContent("@­victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ­ (double-encoded soft hyphen) and neutralize @mention", () => {
+        expect(sanitizeContent("@­victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ‌ (zero-width non-joiner U+200C) and neutralize @mention", () => {
+        expect(sanitizeContent("@‌victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ‍ (zero-width joiner U+200D) and neutralize @mention", () => {
+        expect(sanitizeContent("@‍victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ‎ (left-to-right mark U+200E) and neutralize @mention", () => {
+        expect(sanitizeContent("@‎victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ‏ (right-to-left mark U+200F) and neutralize @mention", () => {
+        expect(sanitizeContent("@‏victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ​ (U+200B) and neutralize @mention", () => {
+        expect(sanitizeContent("@​victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁠ (word joiner U+2060) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁠victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁡ (invisible function application U+2061) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁡victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁡ (U+2061) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁡victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁢ (invisible times U+2062) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁢victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁢ (U+2062) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁢victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁣ (invisible separator U+2063) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁣victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode ⁣ (U+2063) and neutralize @mention", () => {
+        expect(sanitizeContent("@⁣victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode &ip; (invisible plus U+2064) and neutralize @mention", () => {
+        expect(sanitizeContent("@&ip;victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode &InvisiblePlus; (U+2064) and neutralize @mention", () => {
+        expect(sanitizeContent("@&InvisiblePlus;victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode multiple named invisible entities between @ and username", () => {
+        expect(sanitizeContent("@­‌‎victim say hi")).toBe("`@victim` say hi");
+      });
+
+      it("should decode case-insensitive named invisible entities", () => {
+        expect(sanitizeContent("@­victim say hi")).toBe("`@victim` say hi");
+        expect(sanitizeContent("@‌victim say hi")).toBe("`@victim` say hi");
+        expect(sanitizeContent("@‍victim say hi")).toBe("`@victim` say hi");
+        expect(sanitizeContent("@‎victim say hi")).toBe("`@victim` say hi");
+        expect(sanitizeContent("@‏victim say hi")).toBe("`@victim` say hi");
+      });
+    });
   });
 
   describe("template delimiter neutralization (T24)", () => {
diff --git a/actions/setup/js/sanitize_content_core.cjs b/actions/setup/js/sanitize_content_core.cjs
index dc2bacdd199..e66256bcdc4 100644
--- a/actions/setup/js/sanitize_content_core.cjs
+++ b/actions/setup/js/sanitize_content_core.cjs
@@ -966,8 +966,10 @@ function applyTruncation(content, maxLength) {
 /**
  * Decodes HTML entities to prevent bypass of @mention detection and to ensure
  * HTML-encoded characters do not persist in sanitized output (e.g. > in titles).
- * Handles named entities (e.g., @, >, <, &), decimal entities (e.g., @),
- * and hex entities (e.g., @), including double-encoded variants (e.g., @).
+ * Handles named entities (e.g., @, >, <, &, ­, ‌, ‍,
+ * ‎, ‏, ​, ⁠, ⁡/⁡, ⁢/⁢,
+ * ⁣/⁣), decimal entities (e.g., @), and hex entities (e.g., @),
+ * including double-encoded variants (e.g., @).
  *
  * @param {string} text - Input text that may contain HTML entities
  * @returns {string} Text with HTML entities decoded
@@ -993,6 +995,35 @@ function decodeHtmlEntities(text) {
   // & and & → & (decoded after gt/lt so > is already handled above)
   result = result.replace(/&(?:amp;)?amp;/gi, "&");
 
+  // Decode named entities for invisible/formatting characters that are stripped in
+  // hardenUnicodeText Step 3. Without this, the named-entity forms survive entity
+  // decoding and defeat neutralizeAllMentions (e.g. @­user passes the mention
+  // regex because "&" is not in [A-Za-z0-9], then renders as @user in GitHub).
+  // Each entity is decoded to its actual Unicode code point so Step 3 can strip it.
+  // ­ and ­ → U+00AD (soft hyphen)
+  result = result.replace(/&(?:amp;)?shy;/gi, "\u00AD");
+  // ‌ and ‌ → U+200C (zero-width non-joiner)
+  result = result.replace(/&(?:amp;)?zwnj;/gi, "\u200C");
+  // ‍ and ‍ → U+200D (zero-width joiner)
+  result = result.replace(/&(?:amp;)?zwj;/gi, "\u200D");
+  // ‎ and ‎ → U+200E (left-to-right mark)
+  result = result.replace(/&(?:amp;)?lrm;/gi, "\u200E");
+  // ‏ and ‏ → U+200F (right-to-left mark)
+  result = result.replace(/&(?:amp;)?rlm;/gi, "\u200F");
+  // ​ and ​ → U+200B (zero-width space)
+  result = result.replace(/&(?:amp;)?ZeroWidthSpace;/gi, "\u200B");
+  // ⁠ and ⁠ → U+2060 (word joiner)
+  result = result.replace(/&(?:amp;)?NoBreak;/gi, "\u2060");
+  // ⁡ / ⁡ and double-encoded variants → U+2061 (invisible function application)
+  result = result.replace(/&(?:amp;)?(?:af|ApplyFunction);/gi, "\u2061");
+  // ⁢ / ⁢ and double-encoded variants → U+2062 (invisible times)
+  result = result.replace(/&(?:amp;)?(?:it|InvisibleTimes);/gi, "\u2062");
+  // ⁣ / ⁣ and double-encoded variants → U+2063 (invisible separator)
+  result = result.replace(/&(?:amp;)?(?:ic|InvisibleComma);/gi, "\u2063");
+  // &ip; / &InvisiblePlus; and double-encoded variants → U+2064 (invisible plus)
+  // Note: U+2064 is the upper bound of the \u2060-\u2064 range stripped in Step 3.
+  result = result.replace(/&(?:amp;)?(?:ip|InvisiblePlus);/gi, "\u2064");
+
   // Decode decimal entities (including double-encoded variants)
   // @ and @ → @
   // &#NNN; and &#NNN; → corresponding character