By AI Blog Editor
Nov 5, 2026 · 29 min read
The Patch — November 5, 2026
Jan, the desktop app for running models locally, published eleven advisories a week ago, two of them critical, all fixed in 0.8.8. Also today: Mistral Vibe, Pydantic AI, MCP Inspector.
If you chat with local models in Jan, update to 0.8.8: the project published eleven advisories on October 29, two of them critical, and this digest missed them for a week. VulnCheck filed three CVEs last night against Mistral Vibe, Mistral's coding agent, for fixes that shipped in September and October, and Pydantic AI published ten advisories yesterday, fixed in 2.66.0 and 1.107.10.
Component | Affected | Severity | Patched? | Action | Relevance |
|---|---|---|---|---|---|
Jan ×11 | 2 critical (unscored) · 8.8 · 8.1 · 7.8 · 7.6 · 7.3 · 3 medium · low | yes → 0.8.8 (Oct 28) | upgrade; add browser clients to Trusted Hosts | AI stack | |
Mistral Vibe ×3 | 8.8 ×3 (v3.1) · 8.6 (v4) | yes → 3.0.3 (Oct 26) | move off 2.x; 3.1.0 is current | AI stack | |
Pydantic AI ×10 | 7.5 ×2 · 6.8 · 6.5 ×4 · 5.9 · 4.3 · 3.7 | yes → 2.66.0 / 1.107.10 (Nov 4); harness 0.66.0 | upgrade; rotate credentials kept in MCP URLs or embedding headers | AI stack | |
MCP Inspector ×3 | 6.5 · 6.1 · medium | yes → 2.14.0 (Nov 5); 1.0.4 for one | upgrade | AI stack |
None of these is in GitHub's reviewed database. Jan ships as a desktop installer and its advisories name a package GitHub files under "other", so no package scanner will raise them. Mistral Vibe's CVEs are unreviewed records with no package, and Pydantic AI's and the Inspector's are repository advisories without CVEs.
Worth your morning
Jan: eleven advisories, published the day after the fix. Jan (about 45,000 GitHub stars; the 0.8.8 installers have been downloaded about 55,000 times in eight days) runs local models through a bundled llama.cpp engine and can call MCP tools. It shipped 0.8.8 on Wednesday, October 28, calling it a security update, and published eleven advisories on Thursday. Seven affect 0.8.7:
- One-click takeover (GHSA-j6hv, critical, unscored, every release through 0.8.7): content from an MCP server or tool could send Jan's main window to an outside site, which could then use Jan's own commands to start programs. 0.8.8 opens links from model replies, tool results and model cards in the system browser, never inside Jan.
- Claude Code integration (GHSA-959c, 8.8, from 0.7.7): environment variable values set for the integration went into your shell profile unescaped, so a crafted value ran commands in every new shell, and on macOS the elevated fallback could run them as root. 0.8.8 quotes the values and keeps Jan's lines in one marked block. If you used the integration, check your shell profile for lines you didn't add.
- Local API server (GHSA-66pg, 8.1, from 0.6.9): with the server on and no API key, which is the default, any web page you opened could send chat requests to your local model. 0.8.8 checks where browser requests come from.
- File paths (GHSA-9qq3, 7.8, and GHSA-j2wh, medium): unchecked thread and assistant IDs let a caller read, write or delete files outside Jan's data folder.
- Smaller ones: the agent's file tools followed a dangling link out of their workspace (GHSA-7j4g, 4.7, the only one with a CVE, VulnCheck's CVE-2026-122235 from last night), and on Windows Jan ran
explorer.exefrom the search path (low).
The other four cover older lines: binding the API server to 0.0.0.0 in 0.8.0 to 0.8.4 replaced the Trusted Hosts list with a wildcard (GHSA-x6p8, 7.6, fixed in 0.8.5), Model Hub content could run code up to 0.7.5 (critical, fixed in 0.7.6), and two more were fixed in 0.5.15 and 0.6.0.
The upgrade changes behaviour. A browser-based client on another machine, such as Open WebUI's direct connection, or a browser extension calling Jan's API, now needs its host listed under Settings > Local API Server > Trusted Hosts. The CORS switch, which did nothing before, starts on.
There's a second reason to move. Jan 0.8.5 to 0.8.7 bundle llama.cpp build b11146 and run it as llama-server. That build predates b11393, which fixed the 9.2 in llama-server's tool-call parser covered on October 8. 0.8.8 bundles b11669.
Mistral Vibe: three CVEs for fixes the changelog didn't call security. VulnCheck filed three CVEs last night against Mistral Vibe, Mistral's open-source coding agent (about 319,000 PyPI downloads a week). In each, a repository you open runs commands on your machine through Git settings that Vibe's automatic project-context step triggered:
- CVE-2026-122135 (8.8): a signature-verification program set in the repository's Git config, fixed in 3.0.3 (Monday, October 26);
- CVE-2026-122134 (8.8): Git clean or process filters set by the repository, fixed in 2.25.7 (September 22);
- CVE-2026-122133 (8.8, Windows only): a
git.exeplaced in the repository root, fixed in 2.25.5 (September 18).
The 2.x line never got the first fix: 2.26.1, its newest release, doesn't contain the 3.0.3 change. Mistral's changelog lists all three under "Fixed" with no security label, and the repository has no advisories, so the CVEs are the first notice. 3.1.0 (yesterday) is current. Until you're on 3.0.3 or later, don't open repositories you don't trust with Vibe.
Pydantic AI: ten advisories, two that need cleanup after the upgrade. Pydantic AI (about 1.7 million PyPI downloads a week, 8.2 million for pydantic-ai-slim) published ten advisories at 11:51 UTC yesterday, about 80 minutes after 2.66.0 and 1.107.10 reached PyPI. The two highs (7.5) are availability: local web fetching could exhaust memory formatting nested JSON (GHSA-m5vj), and pydantic-ai-harness's private-key redaction could take exponential time on short input (GHSA-mfx5, fixed in harness 0.66.0). Two mediums exposed credentials, and upgrading doesn't take them back:
- 2.57.0 to 2.59.x put a credential carried in an MCP server's URL path into the identifier streamed to chat frontends through MCP Apps events (GHSA-mgmv, 6.5, fixed in 2.60.0 on October 23). Rotate those credentials. The advisory warns that the upgrade also renames the capability's durable operations in Temporal, DBOS or Prefect.
extra_headersandextra_bodyset on an embedder were recorded in traces while instrumentation was on (GHSA-jp94, 6.5). Restrict access to old traces and rotate what they hold.
The rest: the SSRF guard didn't block Tencent Cloud's metadata address (6.8), tool schemas could expand without limit for some models (two 6.5s), tool results shaped like files came back as file downloads when saved history was reloaded (5.9), plus a 4.3 and a 3.7.
Also on the AI stack.
- MCP Inspector 2.14.0 (this morning; about 768,000 npm downloads a week) fixes three mediums in the 2.x web client: an MCP App could get around the content policy the Inspector set for it by navigating its own frame (GHSA-h9j8, 6.1), a malicious server's tool output schema could hang the Inspector (GHSA-2fw8, 6.5, also fixed in 1.0.4 for 1.x), and
server.jsonimports and elicitation URLs went unvalidated (GHSA-rwv7, unscored). - ToolHive 0.62.0 (yesterday) fixes four mediums and a low in Stacklok's MCP server platform and operator. They include Cedar authorization policies that a respelled resource URI or a very large numeric tool argument could get around, and a client
Connectionheader that removed the tokens and headers ToolHive injects. A sixth, a low on fail-open webhooks, was fixed in 0.51.4. - Authlib 1.9.4 (yesterday) fixes a medium in the ID-JAG grant of 1.9.0 to 1.9.3: the documented
authenticate_userresolved users bysubalone, so when a server trusted several identity providers, or a multi-tenant one, one of them could get tokens for another's users (GHSA-9qr6, 5.9). The fix is also a code change on your side: resolve users withissandsub, plustenantfor a multi-tenant provider. It's Authlib's fourth security release in three weeks. - koboldcpp has three highs this digest missed. An unauthenticated generation request with a huge
rep_pen_rangecould run the server out of memory or stall it (GHSA-h59g, 7.5, through 1.122.1). 1.123.1 (October 13) lists the fix. Two more went up on October 31, a quarter of an hour after 1.124.2: a crafted model file in the legacy pre-GGUF formats could write to memory on load and possibly run code (GHSA-4pq2, 7.8, through 1.123.1), and crafted TTS models could crash the loader (GHSA-xv6q, 7.3, 1.98 to 1.122.1). Neither names a fixed version, and 1.124.2's notes don't mention them. Run 1.124.2 and don't load legacy-format or TTS model files from sources you don't trust. - DeepSeek Harness: VulnCheck's CVE-2026-82033 (8.4, last night) covers every build before 0.1.2-alpha.2, including 0.1.2-alpha.1, which carried the August fix covered on October 8: the web profile's HTTP and WebSocket control plane still accepted any local process, including another user's. 0.1.2-alpha.2 (August 30), the build that edition recommended, fixes it. 0.2.1-rc.5 is current.
- Payload 3.93.0 (yesterday) says it "contains vital security updates". Its notes list fixes for block read access in queries, remote-URL upload restrictions, multi-tenant redirects and HTML escaping in the rich-text converters. No advisories have been published yet.
- nginx-ui 2.9.5 (yesterday) fixes OIDC sign-in matching local accounts on untrusted claims and Casdoor tokens accepted from other organizations, under two advisory IDs that aren't public. Setups that matched users by display name or unverified email need a stable
Identifierclaim first. - Ray 2.59.1 (yesterday) brings the eMudhra CA bundle update and aiohttp 3.14.4 to the 2.59 line, for anyone not yet on 2.60.0.
On the Venicecom stack. Nothing new to apply. MSRC opened its November document yesterday with 25 entries, all Microsoft Edge and Chromium. No .NET or Azure entry has been added or revised since yesterday, and Patch Tuesday is November 10. .NET, ASP.NET Core and PrimeNG released nothing, and Angular CLI 22.2.6 (last night) has no security fixes. The MCP C# SDK fix from Tuesday still has no CVE or GitHub database record, so NuGetAudit doesn't flag 1.4.1 or 2.3.0.
Quiet elsewhere. GitHub's reviewed feed published three records on Wednesday; the same query for October 28–29 returns 117, so the feed is working.
Standing items. None of yesterday's repository advisories (Better Auth, SurrealDB, Next.js, Langroid, Authlib's GHSA-gp2m) has reached GitHub's database, and GitHub's MLflow record still names 3.17.2 as the fix for CVE-2026-53995. Better Auth 1.7.17 (yesterday) carries more fixes, among them Expo sign-in now refusing authorization URLs without a state, with no new advisories. SurrealDB has no 3.x release after 3.3.5. LMDeploy's PR 4964 and DeepInfer's PR 242 are still open. AG-UI's three client fixes are still open, so @ag-ui/client 1.0.3 has all three. Prefab UI's advisory still isn't public, and FastMCP 4.1.6 still allows prefab-ui>=0.18.0. Gradio 6.31.0 still requires pillow<13.0. GitHub's oauthlib record 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. kaniko's GHSA-5q7h isn't in GitHub's database, and GO-2026-6225 is unchanged. Open WebUI's promised advisories haven't appeared, and vLLM's three 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.