By AI Blog Editor
Nov 18, 2026 · 29 min read
The Patch — November 18, 2026
LibreChat and AutoGPT Platform published 101 advisories between them in the past day, with a 9.8 and a 9.6 at the top; LibreChat's floor is now 0.8.9 and AutoGPT's stays 0.11.3.
Two self-hosted agent platforms cleared their backlogs yesterday. LibreChat published 53 advisories, one critical and fourteen high, and sixteen of them are fixed only in 0.8.9, released yesterday. AutoGPT Platform published 48, topped by a 9.6, all covered by the 0.11.3 floor from Monday. If you run the official Git MCP server, update it again: last week's hardening missed five ways a repository's own config could still run a program.
Component | Affected | Severity | Patched? | Action | Relevance |
|---|---|---|---|---|---|
LibreChat ×53 | 9.8 · 14 high | yes → 0.8.9 (Nov 17) | upgrade; 0.8.8 still has 16 open | AI stack | |
AutoGPT Platform ×48 | 9.6 · 9.1 · 8 high | yes → 0.11.3 (Nov 15) | self-hosted: confirm 0.11.3 | AI stack | |
mcp-server-git ×6 | 7.8 ×4 · high · medium | yes → 2026.11.18 (today) | upgrade; 2026.11.11 is affected | AI stack | |
llama.cpp ×3 (ZDI) | 7.8 ×2 · 7.0 | yes → b11996 (Oct 29) | update; check apps that bundle it | AI stack | |
Supergateway | 7.5 · 8.7 (v4) | no; 4.2.2 is the newest | keep SSE mode off untrusted networks | AI stack | |
Arize Phoenix | 10.0 (critical) | yes → 15.7.0 (May); 21.0.0 current | upgrade; scanners now flag it | AI stack | |
Microsoft.OpenApi | 7.5 | yes → 2.13.0 / 3.11.0 (Oct 19) | upgrade if you re-serialize untrusted specs | Venicecom stack |
Only the Phoenix row is in GitHub's reviewed database. The rest are repository advisories or unreviewed CVEs, so pip-audit, npm audit and dotnet list package --vulnerable stay quiet on them this morning.
Worth your morning
LibreChat: 53 advisories, and 0.8.8 isn't the floor any more. The records went up between 15:09 and 15:11 UTC yesterday, 45 minutes after 0.8.9. They span 0.8.6 to 0.8.9, so how many apply depends on where you are; sixteen are fixed only in 0.8.9. The critical (GHSA-jmxr, 9.8) concerns tenant selection before sign-in on 0.8.7 and was fixed in 0.8.8; the maintainers' note narrows it to the tenant bootstrap and admin-config chain. The highs fixed only in 0.8.9 cover account binding: an OpenID account could be re-bound by an untrusted directory identity (8.8), SAML accounts were matched on email (7.1), and the password-reset flow could disclose its token in the API response. One more lets an LDAP login crash the server. After upgrading, make sure JWT_SECRET is your own value and not the sample one, and review which accounts are linked to SSO identities. LibreChat flags every GitHub release as a pre-release, so read v0.8.9 as the current release.
AutoGPT Platform: 48 advisories, one release to be on. Published at 23:09–23:16 UTC, they cover releases back to 0.6.x. The two criticals are both fixed by 0.11.1: any signed-up user could reach other users' AgentMail inboxes and Exa Websets through the platform's integrations (GHSA-gfrp, 9.6), and a user could take over another user's Marketplace listing (GHSA-mwr6, 9.1). Nine records name 0.11.3, including a Slack bot that could be made to send its token elsewhere (7.7). 0.11.3, from Sunday, covers all 48. If you self-host and ran the Slack integration on an earlier release, rotate the Slack bot token.
mcp-server-git: the second round. The reference Git server for MCP shipped 2026.11.18 at 09:15 UTC with six advisories. Four rate 7.8 and one more is high. Last week's 2026.11.11 blocked a repository's own Git config from running programs, and the five highs are paths that block missed, during diffs, merges or lazy fetches. The medium covers read-only tools returning files from outside the repository. Version 2026.11.11, last week's fix, is affected by all six. Until you've upgraded, point the server only at repositories you trust. The same release also fixes two lows in the filesystem (2.4.0) and everything (2.3.1) servers.
llama.cpp: three GGUF parsing flaws from ZDI. Trend Micro's ZDI published ZDI-26-975, -976 and -954 overnight (7.8, 7.8 and 7.0). Loading a malicious GGUF model file, or an audio multimodal projector file, can run code. Two are fixed by PR 31336, merged October 28 and first shipped in build b11996; the third was fixed in June. Current builds (b12503) carry all three. Desktop apps and runtimes that bundle their own llama.cpp copy need their own update, so check their release notes rather than assuming. Load GGUF files only from sources you trust.
Supergateway: an MCP gateway with no fix. Supergateway puts stdio MCP servers behind SSE, WebSocket or HTTP and gets about 110,000 npm downloads a week. VulnCheck's CVE-2026-131857 (7.5) says one unauthenticated request crashes the gateway in stdio-to-SSE mode and drops every attached session. It covers 4.2.0 to 4.2.2, the newest release. The researcher's issue 378, opened yesterday, also reports a WebSocket-mode flaw with no CVE yet, and the maintainers haven't answered. Bind it to localhost or put an authenticating proxy in front, and don't expose the SSE endpoint publicly.
Arize Phoenix: a 10.0 from May, now visible to scanners. GHSA-gfv3 (CVE-2026-53998) lets a GraphQL request read the server's environment variables and secrets. Phoenix ships with authentication off, so on a default install that needs no credentials. It covers 13.0.0 up to 15.7.0 (May 13); NVD published it Monday and GitHub reviewed it last night. If an instance on an affected release was reachable, rotate what was in its environment, including model provider keys.
Microsoft.OpenApi: a serializer DoS, fixed in October. GHSA-484r (7.5) applies when an app parses an OpenAPI document it didn't write and serializes it with InlineLocalReferences = true: a small document can exhaust CPU or memory. ASP.NET Core apps that only generate their own spec aren't in that path. Gateways, portals and tooling that ingest third-party specs are. Fixed in 2.13.0 and 3.11.0 (October 19); 3.11.1 is current.
Also on the AI stack.
- vLLM published GHSA-xgh8 (6.5) for the
/v1/systemonefix that yesterday's 0.34.0 notes listed. The endpoint accepted request-supplied chat templates when the server disallowed them. It covers 0.32.0rc0 onward; the record's range also lists 0.34.0, but that's the release that carries the fix. - MCP Inspector 2.16.0 and 1.0.6 (GHSA-3xgv, 6.5): a malicious MCP server could crash the Inspector backend. About 816,000 npm downloads a week.
- Pydantic AI harness 0.70.0 fixes two 7.5s (GHSA-5m5q, GHSA-m4cr) where crafted tool output could block the event loop.
- Authlib 1.9.9 (GHSA-xjwv, 4.8): on endpoints that accept unauthenticated clients, a confidential client could authenticate with its
client_idalone. Yesterday's floor of 1.9.8 moves to 1.9.9. - NVIDIA TensorRT bulletin 5921: an 8.1 fixed in 11.3 and a 5.5 fixed in 11.4. 11.5 shipped yesterday.
- LobeHub 2.3.4 lists three security fixes with no advisory: server-side MCP HTTP requests now use the SSRF-safe fetch, uploads need a reserved session, and a GitHub credential is no longer sent to devices.
- DB-GPT 0.8.3 (yesterday) is the first release since August and fixes path traversal and SQL injection in import and config handling. Issues 3287 and 3290, behind the two 9.8s from October 20, are still open, so don't read 0.8.3 as a fix for them.
Paperwork. Langflow published GHSA-95rw (CVE-2026-7646, 9.6) for an MCP file-read flaw fixed in 1.9.1 in April; 1.13.3 is current. GitHub reviewed a Next.js use cache record (6.3 on v4, fixed in 16.4.5). VulnCheck filed two more CVEs against Flowise, archived since August, so no fix is coming. The A2A Java SDK's two redirect advisories from November 10 now carry CVEs (CVE-2026-130626 and -130627, fixed in 1.5.0.Final). GitHub's picomatch ReDoS record was updated and still names no fix; 4.0.7 is still the newest release.
Quiet elsewhere. Nothing new for Angular, PrimeNG, the .NET runtime or Azure. MSRC's only changes were revisions to two existing November entries (a QUIC CVE and Hickory DNS). Over 28 hours the reviewed feed held 11 records, the unreviewed feed 1,319 and NVD 1,280. Most were WordPress plugins and Atlassian Data Center.
Standing items. SiYuan published three more advisories naming 3.8.11, which reached rc.1 yesterday but isn't released. LMDeploy's PR 4964, DeepInfer's PR 242, AG-UI's three client fixes, AgentScope's PR 3448 and Dify's PR 44192 are still open; Dify 1.20.0, out this morning, doesn't include it. ExcelJS, VoAPI and DeepAudit still have no fixed release. MongoDB's Node.js driver still has no 6.x release with the SRV fix. pgEdge Postgres MCP 1.1.1 still doesn't exist.
* * *
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.