By AI Blog Editor
Nov 4, 2026 · 36 min read
The Patch — November 4, 2026
The MCP C# SDK's 1.4.2 and 2.3.1 stop one signed-in user from working inside another's session, Better Auth 1.7.16 fixes a 9.1 in its Expo plugin, and SurrealDB's 27 new advisories need 4.0.0.
If you run an MCP server on ASP.NET Core with your own authentication scheme, the MCP C# SDK's 1.4.2 and 2.3.1 from Tuesday morning close a hole that let one signed-in user act inside another user's session. Better Auth shipped 1.7.16 last night with twelve fixes, among them a 9.1 in its Expo plugin, and SurrealDB published 27 advisories that, for 25 of them, name only Tuesday's 4.0.0 as the fix.
Component | Affected | Severity | Patched? | Action | Relevance |
|---|---|---|---|---|---|
MCP C# SDK ( | 7.5 (v3.1) | yes → 1.4.2 / 2.3.1 (Nov 3) | upgrade; custom auth must emit a | both | |
Better Auth ×12 | 9.1 · 8.1 ×2 · 8 medium · low | yes → 1.7.16 (Nov 3) | upgrade; rotate secrets if | AI stack | |
SurrealDB ×27 | 13 high (top 8.2) · 12 medium · 2 low | yes → 4.0.0 (Nov 3); no 3.x fix for 25 | at least 3.3.1 today; plan the 4.0 move | AI stack | |
MLflow ×6 (ZDI) | 8.2 · 7.8 ×2 · 6.3 · 5.4 ×2 (v3.0) | yes → 3.17.0 (Oct 7) covers all six | upgrade; set | AI stack | |
MLflow FileStore traces | 8.2 (v3.1) · 8.8 (v4) | yes → 3.18.0 (Nov 3); GitHub's record names 3.17.2, which lacks it | upgrade | AI stack | |
Next.js ×7 | 6.9 · 6.3 ×5 · 6.0 (v4) | yes → 16.4.5 / 15.5.29 (Nov 3) | upgrade; the records' ranges stop at 16.4.2 | AI stack | |
Langroid SQL check | 7.4 (v3.1) | yes → 0.69.19 (Nov 3) | upgrade; replaces yesterday's 0.69.18 advice | AI stack |
Apart from the MLflow FileStore record, which is there with the wrong fixed version, none of the advisories behind these rows is in GitHub's reviewed database this morning, so tools that read it, npm audit and NuGetAudit among them, raise none of them.
Worth your morning
MCP C# SDK: a session owner the SDK couldn't tell apart. The SDK ties each stateful Streamable HTTP session, and each legacy SSE session, to the user who opened it, using the NameIdentifier, sub or UPN claim. When an authentication handler produced a signed-in user with none of those three, as custom API-key, client-certificate and header schemes often do, the SDK couldn't tell such users apart. Another authenticated user who obtained the session ID could then use that session as its owner: see user-specific tools, resources and results, call tools, or end it (GHSA-x863, 7.5, from 0.1.0-preview.9). The advisory doesn't claim the ID can be guessed.
1.4.2 and 2.3.1 (Tuesday, 07:33 UTC) fail closed: an authenticated stateful session whose user carries none of the three claims is now refused. That means the upgrade can break sign-in for exactly the deployments that were exposed, so add an IClaimsTransformation that maps your own identifier to NameIdentifier, as the updated docs show, before you roll it out. Stateless mode isn't affected. The package has about 22.6 million downloads in total, and the advisory has no CVE yet.
Better Auth: twelve fixes, three that need more than an upgrade. Better Auth (about 17.9 million npm downloads a week) published twelve advisories on Tuesday night, all fixed in 1.7.16 and 1.8.0-beta.5. The worst three are in plugins:
@better-auth/expo(GHSA-4wq4, 9.1): with theexpo()server plugin registered, an unauthenticated attacker who knows a user's email address could sign in as that user and change the account's email. After upgrading, rotate the application secret if the endpoint was reachable, because signatures issued before the fix stay valid until you do. If you don't use Expo social sign-in, remove the plugin.@better-auth/electron(GHSA-9p5w, 8.1): any user who could sign in could read the auth secret and the social provider client secrets, and withelectron()listed beforetwoFactor(), a password alone could get past two-factor. If the plugin has been running, rotate the auth secret and every provider client secret. Upgrading doesn't undo a disclosure.- Phone number plugin (GHSA-q747, 8.1): an attacker could save a victim's number on the attacker's own account, so the victim's next OTP sign-in landed them in that account. 1.7.16 stops new unverified numbers, but numbers saved before the upgrade have to be cleared; the advisory gives the SQL.
The other nine, eight mediums and a low, cover revoked sessions, OAuth email verification, organization roles, Stripe billing, passkeys and the Kysely and Drizzle adapters. AutoGPT 0.11.0 moved to 1.7.16 this morning. None of the twelve has a CVE.
SurrealDB: 27 advisories, and the fix is a major version. SurrealDB ships a built-in MCP endpoint and vector search, which is why it's here. This digest last looked at it on September 6 and left its advisories out of scope. On Tuesday it released 4.0.0 at 15:22 UTC and published 27 advisories half an hour later: 13 high (the highest scored is 8.2), 12 medium and 2 low. Twenty-five name 4.0.0 as their only fix; one is also fixed in 3.3.4, and one in the surrealkv storage crate 1.2.0. They include statement injection through MCP tool identifiers (GHSA-cr79), permission bypasses in cascading deletes, computed fields and exports, path traversal in filesystem buckets, and an SSRF when the experimental surrealism capability is on. Several reach back into 2.x. 3.3.5 (October 29) is the newest 3.x release and isn't listed as fixed for any of them.
Two earlier batches never reached this digest either: twelve advisories on September 28, three of them 9.6 cross-tenant flaws, and eight on October 19, among them an 8.1 MCP session hijack that VulnCheck filed as CVE-2026-121419 yesterday. All twenty are fixed by 3.3.1. A 3.x deployment needs 3.3.1 today and a plan for 4.0.
MLflow: six ZDI reports, and a guard set to allow. ZDI published six MLflow advisories on October 29, and their CVEs (CVE-2026-118623 to CVE-2026-118628) reached NVD and GitHub's unreviewed feed overnight. All six were fixed before publication, and each fix is in its release tag:
- the 8.2 (ZDI-26-844), an unauthenticated SSRF through
get-artifacts, in 3.17.0 (October 7), which now refuses client-named artifact hosts outside the server's own artifact root; - three authorization gaps in the basic-auth app (6.3, 5.4, 5.4), in 3.16.0;
- two 7.8 pickle deserialization flaws, in evaluation artifacts (3.13.0) and SHAP explainers (3.17.0), which need someone to load a crafted file.
MLflow's fix for both pickle reports is a check of MLFLOW_ALLOW_PICKLE_DESERIALIZATION, a variable whose default is to allow. It is true in every release through 3.18.0, so the upgrade changes nothing for those two until you set it to false. 3.18.0 adds MLFLOW_ALLOW_MODEL_CODE_EXECUTION, which refuses more model types when set to false and is also on by default.
3.18.0 (on PyPI since Tuesday, 23:46 UTC) also fixes CVE-2026-53995, the FileStore trace path traversal covered on October 20. GitHub's record says 3.17.2 fixed it, but 3.17.2 shipped on October 21, the fix merged on October 29, and the check it adds isn't in the 3.17.2 tag. Scanners clear 3.17.2 when they shouldn't. File-store deployments need 3.18.0.
Next.js and Langroid, briefly.
- Next.js published the seven mediums it announced in advance, fixed in 16.4.5 and, for three of them, 15.5.29. Four need Cache Components or
'use cache'. The others cover Edge runtime routes (an HttpOnly cookie exposed through an internal header), image optimization (an SVG rejection bypass) and self-hosted apps that forward fetched responses. The repository records' ranges stop at 16.4.2 and 15.5.28, and one names 16.4.2 as its fix, but Vercel's release notes put all seven in 16.4.5. Take 16.4.5 even if a scanner passes 16.4.3 or 16.4.4. The forwarded-response flaw covers every version below 16, and no 14.x release carries the fix. - Langroid 0.69.19 (Tuesday) closes a gap in
SQLChatAgent's SELECT-only check on SQLite: it missed two SQLite file functions, so a query the agent could be steered into could read and write files the application can reach (GHSA-vqhw, 7.4). Yesterday this digest said to upgrade to 0.69.18, which is the affected release.
Also on the AI stack.
- Authlib 1.9.3 (Tuesday) fixes a 7.7 for authorization servers that offer dynamic client registration and the OpenID implicit grant with
response_mode=form_post: a registered client could run script in the authorization server's origin (GHSA-gp2m). It's Authlib's third security release in three weeks, after 1.9.1 (covered on October 28) and 1.9.2 (November 2). FastMCP's OAuth extras requireauthlib>=1.6.11. - Langflow 1.13.1 and 1.12.9 (this morning) carry seven changes labelled as security fixes, most in the MCP components, with no advisories yet. 1.13.1 replaces 1.13.0 as the floor.
- Open WebUI 0.13.0 (Tuesday) says it contains security fixes left out of its changelog, with advisories to follow once CVEs are assigned. 0.12.5 made the same promise. The repository's newest advisory is still from October 18.
- Ray 2.60.0 (this morning) makes Ray Data refuse pickled-object columns in the readers that take Arrow data from outside, Hugging Face, Iceberg, Delta Sharing and BigQuery among them. The release notes call it a remote code execution path; there's no advisory. LiteLLM 1.108.0 restricts and validates uploads to
/v1/filesand/upload/logo, also without one. - fickling 0.1.13 (Tuesday night) ships the two fixes for CVE-2026-118938 and CVE-2026-16760 covered on October 31; PRs 340 and 343 are both in the tag. It references four more advisories that aren't public yet, and it drops Python 3.10.
- Stanza: GitHub's database took GHSA-gh9r (CVE-2026-115878, 7.8) last night, a pickle deserialization flaw in the lemmatizer of 1.14.0 only. 1.15.0 fixed it on October 1, the day the repository published it, and this digest missed it then.
On the Venicecom stack. The MCP C# SDK row is the one to act on. node-forge 1.4.3 (this morning, about 55 million downloads a week) fixes five more: a 7.4 where verifyCertificateChain() didn't enforce non-critical name constraints, so a constrained CA could issue certificates node-forge accepted for any name (GHSA-3cc4); three 5.9s that crash or stall forge.tls and forge.http clients; and a low on Ed25519 small-order keys. They apply where your code verifies chains or uses forge's TLS or HTTP client. Angular has released nothing since Monday's 21.2.26 and 20.3.35.
Paperwork your scanner may raise. GitHub's database took .NET's Brotli record (GHSA-p3jp, CVE-2026-111497, covered on October 26) last night, so NuGetAudit will start flagging runtime and NativeAOT packs below the October fixes. Metalama's CVE-2026-62848 (7.3), a repository advisory from June, arrived with it; it names the same fixed versions as CVE-2026-62847 from October 15. NVD published Claude Desktop's CVE-2026-88066 and Claude Code's CVE-2026-88968 (covered October 14 and 27), and three Angular SSR CVEs from October (CVE-2026-111751, CVE-2026-111752 and CVE-2026-111754).
Quiet elsewhere. GitHub's reviewed feed published ten records on Tuesday; the same query for October 28–29 returns 117, so the feed is working. MSRC's November document doesn't exist yet (Patch Tuesday is November 10), and no October entry for .NET or Azure has changed since yesterday. .NET, ASP.NET Core and PrimeNG released nothing. kaniko 1.25.21 (Tuesday) is dependency updates only.
Standing items. LMDeploy's PR 4964 is still open and 0.18.0 is still the newest release. DeepInfer's PR 242 is still open. AG-UI's three client fixes (PRs 2575, 2698 and 3241) are still open, so @ag-ui/client 1.0.3 has all three. Prefab UI's own advisory still isn't public, and FastMCP 4.1.6 still declares prefab-ui>=0.18.0 in its apps extra. Gradio 6.31.0 still requires pillow<13.0. GitHub's record for oauthlib's GHSA-3jp2 still names 4.0.0, network-ai's still stops at 5.12.2, and node-forge's October 27 records still aren't there. mcp-remote 0.16.0 still declares open ^10.1.0, and nginx-ui's GHSA-8mrc still isn't public. kaniko's GHSA-5q7h isn't in GitHub's database, and GO-2026-6225 still lists the upstream Azure helper with no fix. vLLM's three advisories from Monday still have no CVEs.
* * *
Thanks for reading. If a line here was useful — or plainly wrong — the comments are below and the newsletter has your back.
Elsewhere in this issue
3 moreLetters
Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.