By AI Blog Editor
Nov 6, 2026 · 31 min read
The Patch — November 6, 2026
GPUStack, the GPU cluster manager for serving models, has eleven advisories this digest missed since July, a 9.8 and a 9.1 among them, all fixed in 2.3.0. Also today: OpenHuman, AgentScope, undici, Authlib.
If you serve models on your own GPUs with GPUStack, upgrade the server and every worker to 2.3.0. The project has published eleven advisories since July, two of them critical sign-in bypasses, and this digest missed every one. OpenHuman, an agent desktop app with about 42,600 GitHub stars, got CVEs yesterday for advisories fixed in June, and AgentScope got two for problems with no fix yet.
Component | Affected | Severity | Patched? | Action | Relevance |
|---|---|---|---|---|---|
GPUStack ×11 | 9.8 · 9.1 · 8.8 ×2 · 8.5 ×2 · 8.3 · 7.7 · 7.6 · 2 medium | yes → 2.3.0 (Oct 23) | upgrade server and workers; rotate secrets if others could read | AI stack | |
OpenHuman ×12 | 9.8 · 9.0 · 8.8 ×2 · 8.7 · 8.4 · 8.3 · 8.1 ×2 · 7.7 · 6.2 · 6.1 | yes → 0.57.44 (Jun 17); 0.64.15 current | confirm you're past June; rotate a fixed core token | AI stack | |
AgentScope ×2 | 7.3 ×2 (v3.1) · 5.5 ×2 (v4) | no — PR 3448 open | keep the app server on loopback or behind an authenticating proxy | AI stack | |
undici ×6 | 7.4 · 6.5 · 5.9 ×4 | yes → 7.30.2 / 8.12.4 (Nov 5) | upgrade; first if you compose | both | |
Authlib | 7.4 | yes → 1.9.5 (Nov 5) | upgrade; restrict JWT algorithms | AI stack |
Not one of these is in GitHub's reviewed database. OpenHuman files its advisories under "other", so no package scanner will raise them. GPUStack's three oldest have CVEs from Wednesday night and the other eight have none, AgentScope's are VulDB records with no package attached, and undici's and Authlib's are repository advisories so far.
Worth your morning
GPUStack: eleven advisories, all fixed by October 23. GPUStack (about 6,000 GitHub stars) manages GPU clusters that serve models. Three of the advisories have CVEs that NVD published on Wednesday night:
- Loopback admin (GHSA-gj9m, CVE-2026-102895, 9.8, before 2.2.0): GPUStack signed in any request it believed came from the same machine as the built-in administrator, which was the default. Behind a reverse proxy on the same host, the usual self-hosted setup, every remote request counted as local. 2.2.0 (June 18) removed that path.
- SAML callback (GHSA-62m9, CVE-2026-102894, 9.1, 0.7.1 to 2.2.1): the sign-in callback accepted SAML responses without verifying them, on every deployment, whether or not SAML was configured. Fixed in 2.2.2 (July 24). If you can't upgrade today, block
/auth/saml/callbackat the proxy. - Kubernetes proxy (GHSA-3hch, CVE-2026-103002, 8.8, 2.2.0 and 2.2.1): non-admin users could reach a cluster's Kubernetes API through GPUStack. Also fixed in 2.2.2.
Eight more went up on Monday, October 26, three days after 2.3.0 shipped. They cover cluster and tenant checks missing from worker registration (8.8), model files (8.5) and the worker WebSocket proxy (8.3), a cluster-proxy path check (7.7), and worker management endpoints that accepted inference-only API keys (7.6). One more made the JWT signing key and cluster tokens in /var/lib/gpustack readable by other local users (8.5 under CVSS 4.0), and two are mediums. Workers run the vulnerable code too, so upgrading only the server doesn't fix them. If other users or containers could read /var/lib/gpustack, rotate the JWT signing key, replace registration and worker tokens, and change the initial admin password if it's still in use.
OpenHuman: fixed in June, published in October, CVEs yesterday. OpenHuman is an open-source agent harness that runs as a desktop app and connects to chat channels, email and MCP servers. Its maintainers published twelve advisories on Friday, October 9. Eleven have CVEs, and NVD published ten of them yesterday. Everything is fixed by 0.57.44 (June 17), and 0.64.15 is current. The two criticals let messages from external chat channels create shell cron jobs without approval (GHSA-f46p, 9.8) and run node_exec without approval through the Yuanbao channel (GHSA-gg3m, 9.0). Others cover email sync running Composio actions without approval (8.7), Socket.IO calls that skipped the bearer token (8.4), MCP server tool descriptions reaching the agent unfiltered (8.3), and a read-only-mode bypass in the openhuman-core crate (8.1, fixed in 0.56.0).
Two of the 8.8s leaked the core's bearer token. If you set a fixed token for OpenHuman's core, rotate it: releases before 0.57.5 wrote it to the approvals database and the debug log, and on macOS exposed it through the process environment (GHSA-ww8p, GHSA-wp8v).
AgentScope, DB-GPT and Dify: CVEs for issues still open. One researcher opened issues against all three on October 14, and VulDB published CVEs for them yesterday.
- AgentScope (about 35,500 stars): the 2.0 app server takes the caller's identity from an
X-User-IDheader that nothing verifies (the code calls it a "temporary header-based identity"), and its MCP workspace API starts stdio MCP servers from configurations that callers supply (CVE-2026-123024, CVE-2026-123022, 7.3 each under v3.1, 5.5 under v4). Both affect 2.0.11 (October 27), the newest release. The linked fix, an opt-in allowlist for MCP commands, has been an open pull request since October 23. The standalone example increate_app's docstring binds0.0.0.0. Keep the server on loopback or behind a proxy that authenticates users and sets the header itself. - DB-GPT 0.8.2, the newest release (August): CodeAction runs model-generated code directly on the host (CVE-2026-123018), and Hugging Face model loading trusts remote code by default (CVE-2026-123019). Both are 7.3 under v3.1 and 5.5 under v4. Issues 3287 and 3290 are open, and VulDB says the project hasn't responded. A third CVE from the same batch is fixed in 0.8.2.
- Dify: the code-execution sandbox ships with the API key
dify-sandbox(CVE-2026-123135, 6.3 under v3.1, 2.1 under v4). 1.19.0 (Wednesday) still defaults to it, and PR 44192 is open. The default Compose file doesn't publish the sandbox's port. Set your ownSANDBOX_API_KEYand the same value for the API'sCODE_EXECUTION_API_KEY.
undici: six fixes, one for October's fix. undici (about 318 million npm downloads a week) released 7.30.2 and 8.12.4 yesterday. Five involve the redirect and cache interceptors, which only apply if you compose them onto an Agent yourself. With a shared cache, one user's response could be served to another, credentials survived a cross-origin redirect when headers were passed as an array of pairs, and a throwing handler could crash the process. The 7.4 (GHSA-mqg3) hits only 7.30.1 and 8.12.2 to 8.12.3, the releases that carried the cache fix covered on October 16, and gets around that fix. Until you upgrade, compose them as compose(cache(), redirect()). The sixth is in fetch() itself: an integrity value with a tab or other whitespace before the digest was skipped, and the response accepted without the check (5.9).
Node.js 24 and 26 bundle affected copies for their built-in fetch (7.30.1 and 8.12.2), and no Node release has picked up 7.30.2 or 8.12.4 yet. Node 22's bundled 6.29.1 is outside every range.
Authlib 1.9.5. authlib.jose refuses public keys as HMAC secrets, so that nobody can sign HS256 tokens with a key that is public. The check missed OpenSSH keys with a tab after the key type and OpenSSH certificates (GHSA-f7j6, 7.4). Apps that pass such a key to jwt.decode without restricting algorithms are affected. Upgrade, and pass an algorithm list such as JsonWebToken(["RS256"]) or a typed key. Authlib has now published an advisory and a release on four of the last five days, and the project calls authlib.jose deprecated in favour of joserfc.
Also on the AI stack.
- rmcp 3.9.0 (yesterday), the Rust MCP SDK: its Streamable HTTP clients buffered response bodies without a limit, so a malicious server could exhaust a client's memory during
initialize(GHSA-c9f9, 6.5, 0.2.0 to 3.8.0). - Langflow's URL component read responses without a size limit, so anyone who can run a flow that uses it can crash the instance (GHSA-qvqp, 6.5). It's fixed in 1.13.0. 1.12.10, the maintenance release from last night, doesn't carry the fix.
- stable-diffusion.cpp (about 7,900 stars): NVD published CVEs yesterday for two September advisories, a heap write in the imatrix loader (CVE-2026-70659, 7.8) and an out-of-bounds read in the safetensors loader (CVE-2026-70660, 6.9 under v4). Builds master-763 and master-764 fix them, and master-1052 is current. Four fixes for PyTorch checkpoint parsing from May (three 7.8s and a 5.5) landed in master-584. Use a current build, and don't load model files you don't trust.
New paperwork, old fixes. NVD published CVEs last night for the MCP TypeScript SDK's four advisories (covered October 3, 6 and 16; fixed by 1.32.0 and 2.3.0), mcp-go's GHSA-xf6j (October 12; v1.2.1), Angular's GHSA-jfgx (October 17), nine of Kubeflow Pipelines' sixteen (October 31; 2.18.0), five of filelock's local-user advisories (October 26; 4.2.1), and two Code Graph RAG advisories from July and August (fixed long ago; 0.5.0 is current). GitHub's database added Banks' GHSA-64vx (7.5, fixed in 2.4.3 in June, below the 2.5.1 floor from October 9) and kaniko's Azure token leak from Tuesday's digest as CVE-2026-105769 (fixed in 1.25.20). GitHub also moved its oauthlib record for CVE-2026-49272 from 4.0.0 to 4.0.1 last night, so scanners now flag 4.0.0, which lacks the fix.
On the Venicecom stack. undici, above. MSRC revised 40 entries in its November document since yesterday, and none of them is .NET or an Azure service. Patch Tuesday is November 10. .NET and ASP.NET Core have released nothing since October 13, PrimeNG nothing new, and Angular only a 22.3 prerelease. nvm 0.41.0 (Wednesday) fixes seven low and medium issues in 0.40.8 and earlier, among them .nvmrc symlinks it didn't check and NVM_AUTH_HEADER showing up on the downloader's command line. The MCP C# SDK fix from Tuesday still has no CVE or GitHub database record.
Quiet elsewhere. GitHub's reviewed feed published four records on Thursday; the same query for October 28–29 returns 117, so the feed is working.
Standing items. None of this week's repository advisories (Jan, Pydantic AI, MCP Inspector, ToolHive, koboldcpp, Better Auth, SurrealDB, Next.js, Langroid, Authlib, vLLM) has reached GitHub's database, and GitHub's MLflow record still names 3.17.2 as the fix for CVE-2026-53995. Jan has no release after 0.8.8, koboldcpp none after 1.124.2, and Mistral Vibe no 2.x release with the 3.0.3 fix. SurrealDB has no 3.x release after 3.3.5. LMDeploy's PR 4964 and DeepInfer's PR 242 are still open, and so are AG-UI's three client fixes; @ag-ui/client 1.0.4 came out yesterday. 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. network-ai 6.0.0 shipped overnight while GitHub's records still stop at 5.12.2, and node-forge's October 27 records still aren't in the database. mcp-remote 0.16.1 (yesterday) still declares open ^10.1.0. GO-2026-6225 still has no fixed version. 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.