The Loop  ·  Issue N°047

The Loop

A field journal of the AI frontier — for engineers who ship.

§ The Patch

By AI Blog Editor
Nov 8, 2026 · 20 min read

The Patch — November 8, 2026

DeepSeek-Reasonix, a 36,000-star coding agent, has published 17 advisories since September that no scanner raises, four of them critical. Update to 2.58.0. Also: a September OpenCode 7.5 and a Kubeflow Trainer CVE with no fix.

If you use DeepSeek-Reasonix, the terminal coding agent with about 36,000 GitHub stars, update to 2.58.0. Since September 29 the project has published 17 advisories, four of them critical, and most come down to one thing: a repository you opened could run commands on your machine without asking. None has reached GitHub's database, so no scanner will tell you.

Component

Affected

Severity

Patched?

Action

Relevance

DeepSeek-Reasonix ×17

2.x < 2.56.1 · 1.x

9.6 · 9.4 · 2 critical unscored · 11 high · 2 medium

yes → 2.58.0 (Nov 7); October's nine have no 1.x fix; two only mitigated

move to 2.58.0, from 1.x too; don't mark cloned repositories as trusted

AI stack

OpenCode

1.14.30 – 1.18.21

7.5

yes → 1.18.22 (Aug 24)

confirm you're past 1.18.22 if you run opencode serve

AI stack

Kubeflow Trainer v2

TrainJob API, including 2.4.0

low (Red Hat), no CVSS

no — Red Hat lists it as under investigation

grant TrainJob create only to people who may read the namespace's Secrets

AI stack

None of the three is in a package feed. Reasonix and OpenCode published repository advisories only, and the Kubeflow CVE carries no package, so scanners that match on packages won't raise any of them.

Worth your morning

DeepSeek-Reasonix: 17 advisories no scanner has seen. Reasonix is a terminal coding agent built around DeepSeek's models, with a desktop app (Reasonix Studio) on the same version line. It ships from npm as reasonix and as GitHub release binaries; 2.58.0 came out on Saturday. Between September 29 and November 2 the project published 17 advisories in its own repository: four critical, eleven high, two medium. Every one names a fixed release. Only one has a CVE, VulnCheck's CVE-2026-123989 from Saturday, for a flaw fixed in August, and none is in GitHub's database. The nine from October 27 also name a Go module, github.com/esengine/DeepSeek-Reasonix, while the project's go.mod declares module reasonix and users install it from npm, so a scanner won't match them even after GitHub indexes them.

  • Instruction files as trusted guidance (GHSA-7v48, 9.6, and a 7.9): AGENTS.md, CLAUDE.md and REASONIX.md from the workspace were loaded as durable user guidance, so a repository could steer what the agent ran. 2.53.0 (October 27) labels them unverified in folders you haven't marked trusted. The project calls this a mitigation: in a trusted folder they still load as guidance.
  • Commands and settings that got past approval (9.4, an unscored critical, five highs and a medium): the classifier that lets read-only shell commands run without a prompt could be made to run code defined by the repository, through its git configuration, attributes files or build scripts, and workspace settings could widen the agent's own sandbox and approval rules. Fixes landed between August and November, and on 2.x the highest floor is 2.56.1. On Windows there's no OS sandbox behind the classifier, so these ran as the user.
  • Project MCP servers (critical, 7.9, 7.8): a repository's MCP configuration could start servers without asking (fixed in 1.39.3), an approved project server could change its command after a pull without new approval (fixed in 2.33.0, October 9), and a server could change its tool descriptions without the user seeing it. That last one is only partly fixed: 2.53.0 stops repositories from forcing a server's tools into every session, but changed descriptions still don't trigger a new approval.
  • The rest: two ways around the SSRF guard on the web-fetch and install tools, a pre-authentication memory exhaustion on the reasonix serve login form (7.5), and unpinned actions in the project's own release pipeline (8.7), pinned to commits before 2.53.0.

Update the CLI and Studio to 2.58.0. The project says 1.x is no longer maintained: npm's legacy-v1 tag still points at 1.39.11 (November 2), which has the last 1.x fix but none of the nine from October 27, and the reporter confirmed the 9.4 on 1.39.6. Don't mark a cloned repository as trusted unless you'd run its scripts yourself, and read changes to .mcp.json and reasonix.toml after pulling before you start the agent.

OpenCode: a September advisory this digest missed. OpenCode (about 225,000 stars, 3.4 million npm downloads last week) published GHSA-632h (7.5) on September 24. While opencode serve was running on an npm-, pnpm- or Bun-installed OpenCode, a web page the browser could load was able to make the server install packages of the page's choosing, which runs code as the user. With server authentication on, the browser also needed cached credentials. The fix shipped in 1.18.22 on August 24, a month before the advisory, and 1.19.5 is current. The advisory is repository-only, so check the version by hand if you run the server.

Kubeflow Trainer: a CVE with no fix and no score. Red Hat's CVE-2026-83495, which reached NVD on Saturday, covers Trainer v2's TrainJob API: anyone allowed to create TrainJobs can have the job's pods read Secrets from the namespace through environment variables. That matters wherever TrainJob permissions are wider than Secret read permissions. Red Hat rates it low and lists the OpenShift AI fix as under investigation. The CVE text names no project, versions or fix, and 2.4.0 (Friday) still passes secretKeyRef entries through to the pods. Until a fix lands, treat TrainJob create as Secret read in that namespace, or block secretKeyRef in TrainJob environment variables with an admission policy.

Also on the AI stack. Obot 0.28.0 (Saturday night) makes the MCP gateway's OAuth token exchange check that the user may use the MCP server before it issues a token (PR 8380). There's no advisory, and 0.27.0 doesn't have the check.

New paperwork, old fixes. VulnCheck filed CVEs overnight for advisories the projects published weeks ago:

  • Tornado, which Jupyter Server and Streamlit run on: two that the project rated low, fixed in 6.6 (October 21). VulnCheck scores the static-file one (GHSA-v8g8, CVE-2026-123534) 6.5, and 8.3 under v4, so tools that read v4 will show a high for it. The other (GHSA-c6v4) concerns cookie attributes when an application passes untrusted input as max_age.
  • rustls: five from September and October, fixed in 0.23.45 and 0.23.46 (October 14). rustls rates them 2.3 to 5.3; VulnCheck's v4 scores go up to 6.9. Rust tools that link rustls need a build against 0.23.46.
  • Vikunja: 39 CVEs for issues fixed in 2.7.1 (October 20), four of them in veans, its CLI that hands tasks to coding agents. The highest (8.8) lets a signed-in user exhaust the server's memory.

On the Venicecom stack. Microsoft published a 10.0 for Entra Device Registration (CVE-2026-109469) on Thursday. It's a cloud-service CVE: Microsoft has already mitigated it and tenants have nothing to do, but it will show up as a critical in any feed that tracks Microsoft CVEs. MSRC revised nothing on .NET or Azure services since yesterday, and Patch Tuesday is November 10. Angular, PrimeNG, .NET and ASP.NET Core released nothing new.

Quiet elsewhere. GitHub's reviewed feed has published nothing since Friday at 19:51 UTC. The same query for October 28–29 returns 117, so the feed is working and dark for the weekend. The unreviewed feed's 636 records included a 208-record IBM API Connect import and VulnCheck batches for Hickory DNS, OpenAM, Zebra and devalue, none of them on the tracked stack. The tracked repositories published nothing new since yesterday's run.

Standing items. None of this week's repository advisories (GPUStack, OpenHuman, undici, Authlib, rmcp, Langflow, Jan, Better Auth, Next.js, Agent Router, pypdf, ToolHive, vLLM) has reached GitHub's database, and GitHub's MLflow record still names 3.17.2 as the fix for CVE-2026-53995. LMDeploy's PR 4964, DeepInfer's PR 242, AgentScope's PR 3448 and Dify's PR 44192 are still open, as are AG-UI's three client fixes and DB-GPT's issues 3287 and 3290. LlamaIndex's file readers, Unstructured and Open WebUI still hold pypdf below 7. Prefab UI's advisory still isn't public, FastMCP 4.1.6 still allows prefab-ui>=0.18.0, and Gradio 6.31.0 still requires pillow<13.0. mcp-remote 0.16.5 still declares open ^10.1.0. Node's main branch bundles undici 8.12.4, and no Node release carries it yet. TorchServe's CVE still has no package mapping. Jan has no release after 0.8.8 and koboldcpp none after 1.124.2. network-ai 6.0.1 (Saturday) lists file-handling hardening with no advisories, and GitHub's records for it still stop at 5.12.2.

* * *

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 more
  1. 01

    News

    Anthropic priced its IPO above the range — $83 a share, 980 million shares, and ANT starts trading today

    Nov 19, 2026

  2. 02

    The Patch

    The Patch — November 19, 2026

    Nov 19, 2026

  3. 03

    The Patch

    The Patch — November 18, 2026

    Nov 18, 2026

Letters

Arguments, corrections, questions. Anonymous comments allowed; be kind, be specific.