diff --git a/CHANGELOG.md b/CHANGELOG.md
index 17cda04..c9fd451 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,73 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [10.3.1] - 2026-09-10
+
+Fixes a polynomial ReDoS reachable from attacker-controlled input in the HTML
+sanitizer, the template renderer and the router. No API change.
+
+### Security
+- **`HtmlSanitizer` could be stalled for minutes by one request (high).** Every
+  tag pattern ended in `[^>]*`. On input containing many `<` and no `>`, the
+  regex engine rescanned to the end of input from each `<` and backtracked,
+  making the work quadratic in input size. `MAX_INPUT_SIZE` did not help — 1MB
+  is *inside* the cap, and quadratic growth at that size extrapolated to roughly
+  19 minutes of blocked event loop. Node is single-threaded, so a single request
+  stalled the whole process; on Lambda it is a timeout you are billed for.
+
+  Measured before the fix: 20KB took 421ms, 40KB 1.6s, 80KB 6.6s — a clean 4x
+  per doubling. After: 500KB completes in under 60ms.
+
+  The tag patterns now match unquoted characters and fully quoted attribute
+  values separately. The alternatives are disjoint on their first character, so
+  the match is deterministic and linear. A naive `[^<>]*` would also have been
+  linear but wrong — a quoted attribute value may legally contain `<`, and
+  ending the tag there let attribute content escape into the document as
+  structure.
+
+  CodeQL flagged two of these sites; the same defect was present at three more
+  that it did not report. All five are fixed.
+
+- **`Router` path normalization was quadratic (moderate).** `path.replace(/\/+$/, '')`
+  has an anchored `+`, so the engine retried from every index in a run of
+  slashes, scanning to the end each time. Request paths are attacker-controlled,
+  though bounded by HTTP header limits, so the practical cost was ~100-260ms of
+  CPU per request rather than minutes. Trailing slashes are now stripped with an
+  index scan.
+
+- **`HtmlRenderer` `{{#if}}` condition matching was quadratic (low).** `\s+([^}]+)`
+  — both parts match a space, so the engine tried every split point and rescanned
+  to the end for each. Exploiting it requires an attacker-controlled *template*;
+  values passed as data go through the slot table and are never re-parsed as
+  template syntax, so ordinary use is not affected. `HtmlRenderer` has no input
+  cap, so it is fixed regardless: the condition must now start with a non-space,
+  making the two parts disjoint.
+
+### Changed
+- Sanitizing malformed input where a stray `<` immediately precedes a tag now
+  removes the whole construct instead of leaving part of it behind.
+  `HtmlSanitizer.clean('<<b>>')` returns `''` where it previously returned
+  `'&gt;'`. This is the only observable output difference — verified by diffing
+  every sanitizer and renderer output against 10.3.0 across a payload corpus —
+  and it errs toward removing more, not less.
+
+### Added
+- ReDoS regression tests for all three modules, asserting linear-time completion
+  on pathological input. Against the unfixed code these take 4-55 seconds each
+  and fail; against the fix they complete in milliseconds, so the bounds are not
+  timing-sensitive.
+- Sanitizer bypass tests covering multi-character sanitization: nested tag
+  reassembly (`<scr<script>ipt>`), triple nesting, comment splitting, entity and
+  hex encoding, tab-broken `javascript:` URLs, and `on*` handlers smuggled
+  through attribute values.
+
+  These pin a mitigation CodeQL cannot see. It reports
+  `js/incomplete-multi-character-sanitization` on the individual `.replace()`
+  calls, but the sanitizer repeats the strip until output is stable, escapes any
+  residual `<`, and re-emits only allowlisted tags from a builder. Those two
+  alerts are dismissed as false positives; these tests are what keeps that
+  dismissal honest, by failing if the mitigation is ever removed.
+
 ## [10.3.0] - 2026-09-10
 
 Type-definition fixes found while integrating the package into an AWS Lambda +
diff --git a/package-lock.json b/package-lock.json
index a241850..1748893 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "@designofadecade/server",
-  "version": "10.3.0",
+  "version": "10.3.1",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@designofadecade/server",
-      "version": "10.3.0",
+      "version": "10.3.1",
       "license": "MIT",
       "dependencies": {
         "ws": "^8.21.3"
diff --git a/package.json b/package.json
index 7239dda..430021c 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@designofadecade/server",
-  "version": "10.3.0",
+  "version": "10.3.1",
   "description": "Design of a Decade Server - A modern TypeScript server with WebSocket support",
   "type": "module",
   "main": "./dist/index.js",
diff --git a/src/router/Router.security.test.ts b/src/router/Router.security.test.ts
new file mode 100644
index 0000000..7293cb6
--- /dev/null
+++ b/src/router/Router.security.test.ts
@@ -0,0 +1,50 @@
+import { describe, it, expect } from 'vitest';
+import Router from './Router.ts';
+import Routes from './Routes.ts';
+
+/**
+ * Route paths are attacker-controlled. Trailing slashes were stripped with
+ * `replace(/\/+$/, '')`, whose anchored `+` makes the engine retry from every
+ * index in a run of slashes, scanning to the end each time — quadratic. A
+ * request of many slashes turned into O(n^2) work on the shared event loop.
+ */
+describe('Router path normalization ReDoS resistance', () => {
+  const BUDGET_MS = 2000;
+
+  const build = () =>
+    new Router({
+      initRoutes: [
+        class extends Routes {
+          constructor(router: Router) {
+            super(router);
+            this.addRoute('/ok', ['GET'], async () => ({ status: 200, body: 'ok' }));
+          }
+        },
+      ],
+    });
+
+  it('normalizes a long run of slashes in linear time', async () => {
+    const router = build();
+    const started = performance.now();
+    await router.lambdaEvent({
+      requestContext: { http: { method: 'GET', path: '/'.repeat(200_000) + 'x' } },
+      headers: {},
+    });
+    expect(performance.now() - started).toBeLessThan(BUDGET_MS);
+  });
+
+  it('still routes and still strips trailing slashes', async () => {
+    const router = build();
+    const hit = await router.lambdaEvent({
+      requestContext: { http: { method: 'GET', path: '/ok///' } },
+      headers: {},
+    });
+    expect(hit.statusCode).toBe(200);
+
+    const root = await router.lambdaEvent({
+      requestContext: { http: { method: 'GET', path: '/' } },
+      headers: {},
+    });
+    expect(root.statusCode).toBe(404);
+  });
+});
diff --git a/src/router/Router.ts b/src/router/Router.ts
index e554746..7237552 100644
--- a/src/router/Router.ts
+++ b/src/router/Router.ts
@@ -180,8 +180,7 @@ export default class Router {
 
       // Normalize: collapse multiple slashes, remove trailing slash (except root)
       let normalizedPath = route.path.replace(/\/+/g, '/');
-      normalizedPath =
-        normalizedPath.length > 1 ? normalizedPath.replace(/\/+$/, '') : normalizedPath;
+      normalizedPath = Router.#stripTrailingSlashes(normalizedPath);
 
       for (const method of route.methods) {
         const pathMethodKey = `${normalizedPath}::${method}`;
@@ -214,7 +213,7 @@ export default class Router {
     method: string
   ): RouteRegistration | { handler: (request: RouterRequest) => Promise<RouterResponse> } | null {
     // Normalize path: remove trailing slash (except for root "/")
-    const normalizedPath = path.length > 1 ? path.replace(/\/+$/, '') : path;
+    const normalizedPath = Router.#stripTrailingSlashes(path);
 
     if (this.#routes.static.has(`${normalizedPath}::${method}`))
       return this.#routes.static.get(`${normalizedPath}::${method}`)!;
@@ -236,6 +235,20 @@ export default class Router {
     return route || null;
   }
 
+  /**
+   * Strip trailing slashes, keeping a lone "/" intact.
+   *
+   * `replace(/\/+$/, '')` looks equivalent but is quadratic: the anchored `+`
+   * makes the engine retry from every index in a run of slashes, each time
+   * scanning to the end before failing. A path is attacker-controlled, so a
+   * request of many slashes turned into O(n^2) work on the event loop.
+   */
+  static #stripTrailingSlashes(path: string): string {
+    let end = path.length;
+    while (end > 1 && path.charCodeAt(end - 1) === 47 /* '/' */) end--;
+    return end === path.length ? path : path.slice(0, end);
+  }
+
   #pruneCache(): void {
     if (this.#routes.cache.size > Router.#MAX_CACHE_SIZE) {
       const firstKey = this.#routes.cache.keys().next().value;
diff --git a/src/sanitizer/HtmlSanitizer.security.test.ts b/src/sanitizer/HtmlSanitizer.security.test.ts
new file mode 100644
index 0000000..f2632f7
--- /dev/null
+++ b/src/sanitizer/HtmlSanitizer.security.test.ts
@@ -0,0 +1,104 @@
+import { describe, it, expect } from 'vitest';
+import HtmlSanitizer from './HtmlSanitizer.ts';
+
+/**
+ * Sanitizer input is attacker-controlled by definition, so the cost of
+ * sanitizing has to stay linear in input size.
+ *
+ * The tag patterns used to end in `[^>]*`. On input with many '<' and no '>'
+ * the engine rescanned to the end of input from every '<' and backtracked,
+ * making the work quadratic: 80KB took ~6.6s and 1MB — inside MAX_INPUT_SIZE —
+ * extrapolated to roughly 19 minutes of blocked event loop for one request.
+ *
+ * The bounds below are deliberately loose. Post-fix these inputs take tens of
+ * milliseconds, so a 2s ceiling is ~100x headroom against CI jitter while still
+ * failing decisively if the quadratic behaviour returns — against the unfixed
+ * code these same cases take 4-55 seconds each.
+ *
+ * These assert an absolute budget rather than a growth ratio between two input
+ * sizes. At millisecond scale a ratio is dominated by JIT warmup, which made it
+ * flaky; the absolute ceiling has orders of magnitude more headroom.
+ */
+describe('HtmlSanitizer ReDoS resistance', () => {
+  const BUDGET_MS = 2000;
+  const N = 200_000;
+
+  const timed = (fn: () => unknown): number => {
+    const started = performance.now();
+    fn();
+    return performance.now() - started;
+  };
+
+  it('sanitizes a long run of "<" in linear time', () => {
+    expect(timed(() => HtmlSanitizer.clean('<'.repeat(N)))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('sanitizes repeated dangerous-tag prefixes in linear time', () => {
+    const payload = '<link' + '<base'.repeat(N / 5);
+    expect(timed(() => HtmlSanitizer.clean(payload))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('sanitizes unterminated tags in linear time with an allowlist', () => {
+    const payload = '<a href="'.repeat(N / 9);
+    expect(timed(() => HtmlSanitizer.clean(payload, ['a', 'b']))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('strips tags from long unbalanced input in linear time', () => {
+    const payload = '<b>ok</b>' + '<'.repeat(N);
+    expect(timed(() => HtmlSanitizer.clean(payload, ['b']))).toBeLessThan(BUDGET_MS);
+  });
+});
+
+/**
+ * CodeQL reports `js/incomplete-multi-character-sanitization` on the individual
+ * tag-stripping `.replace()` calls, because a single pass over `<scr<script>ipt>`
+ * reassembles `<script>`. That is mitigated structurally rather than locally:
+ * the strip loop repeats until the output stops changing, any residual '<' is
+ * escaped, and output tags are re-emitted from an allowlist builder.
+ *
+ * Those alerts are dismissed as false positives. This suite is what makes that
+ * dismissal safe — remove any part of the mitigation and these fail, rather
+ * than the hole reopening silently.
+ */
+describe('HtmlSanitizer multi-character sanitization', () => {
+  const executable = (s: string) =>
+    /<\s*(script|iframe|object|embed|applet)|\son\w+\s*=|javascript\s*:/i.test(s);
+
+  const payloads = [
+    '<scr<script>ipt>alert(1)</scr</script>ipt>',
+    '<scri<script>pt>alert(1)</scri</script>pt>',
+    '<<script>script>alert(1)<</script>/script>',
+    '<scr<scr<script>ipt>ipt>alert(1)</scr</scr</script>ipt>ipt>',
+    '<sc<!-- -->ript>alert(1)</sc<!-- -->ript>',
+    '<img sr<script>c=x onerror=alert(1)>',
+    '<img src=x on<script>error=alert(1)>',
+    '<a hr<script>ef="javascript:alert(1)">x</a>',
+    '<a href="jav&#x61;script:alert(1)">x</a>',
+    '<a href="java\tscript:alert(1)">x</a>',
+    '<a href="&#106;avascript:alert(1)">x</a>',
+    '<b onmouseover=alert(1)>hi</b>',
+    '<iframe sr<script>c="//evil">',
+    '&lt;script&gt;alert(1)&lt;/script&gt;',
+    '<base<base href="//evil">href="//evil">',
+  ];
+
+  for (const payload of payloads) {
+    it(`never emits executable markup for ${JSON.stringify(payload).slice(0, 44)}`, () => {
+      expect(executable(HtmlSanitizer.clean(payload))).toBe(false);
+      expect(executable(HtmlSanitizer.clean(payload, ['a', 'b', 'i', 'img']))).toBe(false);
+    });
+  }
+
+  /**
+   * A quoted attribute value may legally contain '<'. The linear tag pattern
+   * has to consume the whole quoted value, or attribute content escapes into
+   * the document as structure.
+   */
+  it('does not leak quoted attribute content as markup', () => {
+    expect(HtmlSanitizer.clean('<a title="a<b">text</a>', ['a', 'b'])).toBe('<a>text</a>');
+    expect(HtmlSanitizer.clean('<a href="/x" title="q<r">t</a>', ['a'])).toBe('<a href="/x">t</a>');
+    expect(HtmlSanitizer.clean("<a title='x<script>alert(1)</script>'>t</a>", ['a'])).toBe(
+      '<a>t</a>'
+    );
+  });
+});
diff --git a/src/sanitizer/HtmlSanitizer.ts b/src/sanitizer/HtmlSanitizer.ts
index 6d3b4d9..40ed27a 100644
--- a/src/sanitizer/HtmlSanitizer.ts
+++ b/src/sanitizer/HtmlSanitizer.ts
@@ -50,6 +50,22 @@ export interface SanitizationResult {
 /**
  * Maximum input size to prevent DoS attacks (1MB)
  */
+/**
+ * Attribute section of an HTML tag.
+ *
+ * A quoted attribute value may legally contain '<', so `[^>]*` cannot simply
+ * become `[^<>]*` — that would end the tag early and let attribute content leak
+ * out as document structure. This matches unquoted characters and fully quoted
+ * values separately. The three alternatives are disjoint on their first
+ * character, so the match is deterministic and runs in linear time.
+ *
+ * SECURITY: the previous `[^>]*` was a polynomial ReDoS. On input with many
+ * '<' and no '>', the engine rescanned to end of input from every '<' and
+ * backtracked, so sanitizing 1MB of attacker-supplied text — well inside
+ * MAX_INPUT_SIZE — blocked the event loop for minutes.
+ */
+const TAG_ATTRS = `(?:[^<>"']|"[^"]*"|'[^']*')*`;
+
 const MAX_INPUT_SIZE: number = 1024 * 1024;
 
 /**
@@ -164,7 +180,7 @@ export default class HtmlSanitizer {
    * region it has already passed.
    */
   static #removeTagBlocks(input: string, tagNames: readonly string[]): string {
-    const opener = new RegExp(`<(${tagNames.join('|')})\\b[^>]*>`, 'gi');
+    const opener = new RegExp(`<(${tagNames.join('|')})\\b${TAG_ATTRS}>`, 'gi');
     const lowered = input.toLowerCase();
 
     // Once a closing tag is absent from the remainder it is absent for every
@@ -397,7 +413,10 @@ export default class HtmlSanitizer {
 
     // Also remove self-closing dangerous tags
     cleaned = cleaned.replace(
-      /<(script|style|iframe|object|embed|applet|link|base|meta|form|input|button|textarea|select)[^>]*\/?>/gi,
+      new RegExp(
+        `<(script|style|iframe|object|embed|applet|link|base|meta|form|input|button|textarea|select)${TAG_ATTRS}\\/?>`,
+        'gi'
+      ),
       ''
     );
 
@@ -564,7 +583,7 @@ export default class HtmlSanitizer {
       return `<${lowerTagName}>`;
     };
 
-    const tagPattern = /<\/?([a-z][a-z0-9]*)\b[^>]*>/gi;
+    const tagPattern = new RegExp(`</?([a-z][a-z0-9]*)\\b${TAG_ATTRS}>`, 'gi');
     let rebuilt = '';
     let cursor = 0;
     let tagMatch: RegExpExecArray | null;
@@ -664,7 +683,7 @@ export default class HtmlSanitizer {
 
     // Replace opening tags of the specified type
     return html.replace(
-      new RegExp(`<${lowerTagName}\\b([^>]*)>`, 'gi'),
+      new RegExp(`<${lowerTagName}\\b(${TAG_ATTRS})>`, 'gi'),
       (_match: string, existingAttrs: string): string => {
         // Parse existing attributes to preserve href, src, etc.
         const preservedAttrs: string[] = [];
@@ -737,7 +756,7 @@ export default class HtmlSanitizer {
       previousStripped = text;
       text = HtmlSanitizer.#removeComments(text);
       text = HtmlSanitizer.#removeTagBlocks(text, ['script', 'style']);
-      text = text.replace(/<[^>]*>/g, '');
+      text = text.replace(new RegExp(`<${TAG_ATTRS}>`, 'g'), '');
       stripPass++;
     }
 
diff --git a/src/utils/HtmlRenderer.security.test.ts b/src/utils/HtmlRenderer.security.test.ts
new file mode 100644
index 0000000..6daa22c
--- /dev/null
+++ b/src/utils/HtmlRenderer.security.test.ts
@@ -0,0 +1,81 @@
+import { describe, it, expect } from 'vitest';
+import HtmlRenderer from './HtmlRenderer.ts';
+
+/**
+ * `{{#if}}` matched its condition with `\s+([^}]+)`. Both parts match a space,
+ * so on a template like `{{#if` followed by a long run of spaces and no closing
+ * `}}`, the engine tried every split point and rescanned to end of input for
+ * each — quadratic. Requiring the condition to start with a non-space makes the
+ * two disjoint, so a failed match backtracks in constant time per step.
+ *
+ * A condition also cannot contain braces, which bounds each scan at the next
+ * `{` rather than at end of input — without that, a repeated `{{#if` gave many
+ * start positions that each rescanned the whole template.
+ *
+ * HtmlRenderer has no input cap, so this is bounded only by template size.
+ */
+describe('HtmlRenderer ReDoS resistance', () => {
+  const BUDGET_MS = 1000;
+
+  const timed = (fn: () => unknown): number => {
+    const started = performance.now();
+    fn();
+    return performance.now() - started;
+  };
+
+  it('handles an unterminated {{#if}} with a long whitespace run', () => {
+    const template = '{{{{#if ' + ' '.repeat(100_000);
+    expect(timed(() => HtmlRenderer.render(template, {}))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('handles an unterminated {{#if}}/{{else}} with a long whitespace run', () => {
+    const template = '{{{{#if ' + ' '.repeat(50_000) + '{{else}}';
+    expect(timed(() => HtmlRenderer.render(template, {}))).toBeLessThan(BUDGET_MS);
+  });
+
+  /**
+   * The first attempt at this fix only stopped the condition from starting with
+   * whitespace, which fixed a single long run but not the general case: with
+   * `{{#if` repeated, every occurrence is a start position whose condition scan
+   * still ran to end of input. Excluding braces from the condition bounds each
+   * scan at the next `{`, which is what makes it linear.
+   */
+  it('handles many repeated unterminated {{#if}} openings', () => {
+    const template = '{{{{#if !'.repeat(32_000);
+    expect(timed(() => HtmlRenderer.render(template, {}))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('handles many {{#if}} openings with no matching close', () => {
+    const template = '{{#if a}}'.repeat(32_000);
+    expect(timed(() => HtmlRenderer.render(template, { a: 1 }))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('handles many {{#if}}/{{else}} openings with no matching close', () => {
+    const template = '{{#if a}}x{{else}}'.repeat(32_000);
+    expect(timed(() => HtmlRenderer.render(template, { a: 1 }))).toBeLessThan(BUDGET_MS);
+  });
+
+  it('handles many repeated unterminated {{#each}} openings', () => {
+    const template = '{{#each a}}'.repeat(32_000);
+    expect(timed(() => HtmlRenderer.render(template, { a: [] }))).toBeLessThan(BUDGET_MS);
+  });
+
+  /**
+   * The condition capture changed from `[^}]+` to `[^}\s][^}]*`, so it must
+   * still accept every condition form the renderer supports — including the
+   * comparison operators, whose `<` and `>` the capture has to pass through,
+   * and surrounding whitespace, which `#evaluateCondition` trims.
+   */
+  it('still renders every supported condition form', () => {
+    expect(HtmlRenderer.render('{{#if a}}Y{{/if}}', { a: true })).toBe('Y');
+    expect(HtmlRenderer.render('{{#if a}}Y{{else}}N{{/if}}', { a: false })).toBe('N');
+    expect(HtmlRenderer.render('{{#if  a  }}Y{{/if}}', { a: true })).toBe('Y');
+    expect(HtmlRenderer.render('{{#if a }}Y{{/if}}', { a: true })).toBe('Y');
+    expect(HtmlRenderer.render('{{#if a.b}}Y{{/if}}', { a: { b: 1 } })).toBe('Y');
+    expect(HtmlRenderer.render('{{#if a === 1}}Y{{/if}}', { a: 1 })).toBe('Y');
+    expect(HtmlRenderer.render('{{#if a !== 1}}Y{{else}}N{{/if}}', { a: 1 })).toBe('N');
+    expect(HtmlRenderer.render('{{#if a > 2}}Y{{/if}}', { a: 3 })).toBe('Y');
+    expect(HtmlRenderer.render('{{#if a < 2}}Y{{else}}N{{/if}}', { a: 3 })).toBe('N');
+    expect(HtmlRenderer.render('{{#if a >= 2}}Y{{/if}}', { a: 2 })).toBe('Y');
+  });
+});
diff --git a/src/utils/HtmlRenderer.ts b/src/utils/HtmlRenderer.ts
index bdf89a8..8210ddf 100644
--- a/src/utils/HtmlRenderer.ts
+++ b/src/utils/HtmlRenderer.ts
@@ -91,7 +91,7 @@ export default class HtmlRenderer {
 
       // Process {{#if}} with {{else}} - innermost first (non-greedy match)
       processed = processed.replace(
-        /\{\{#if\s+([^}]+)\}\}((?:(?!\{\{#if|\{\{#each).)*?)\{\{else\}\}((?:(?!\{\{#if|\{\{#each).)*?)\{\{\/if\}\}/s,
+        /\{\{#if\s+([^{}\s][^{}]*)\}\}((?:(?!\{\{#if|\{\{#each).)*?)\{\{else\}\}((?:(?!\{\{#if|\{\{#each).)*?)\{\{\/if\}\}/s,
         (_match, condition, ifContent, elseContent) => {
           const isTruthy = HtmlRenderer.#evaluateCondition(condition, vars);
           return isTruthy ? ifContent : elseContent;
@@ -100,7 +100,7 @@ export default class HtmlRenderer {
 
       // Process {{#if}} without {{else}} - innermost first (non-greedy match)
       processed = processed.replace(
-        /\{\{#if\s+([^}]+)\}\}((?:(?!\{\{#if|\{\{#each).)*?)\{\{\/if\}\}/s,
+        /\{\{#if\s+([^{}\s][^{}]*)\}\}((?:(?!\{\{#if|\{\{#each).)*?)\{\{\/if\}\}/s,
         (_match, condition, content) => {
           const isTruthy = HtmlRenderer.#evaluateCondition(condition, vars);
           return isTruthy ? content : '';