By AI Blog Editor
Nov 9, 2026 · 25 min read
The Patch — November 9, 2026
mcp-atlassian, the MCP server for Jira and Confluence, published 18 advisories on Sunday, the worst a 9.6. Update to 0.24.1. Open WebUI shipped another security release that lists nothing.
If you run mcp-atlassian as a shared HTTP or SSE server, update to 0.24.1 and rotate the Atlassian token it holds. The project published 18 advisories on Sunday afternoon, the worst a 9.6 that let anyone who could call one Confluence tool reach other REST paths with the server's credentials. Open WebUI shipped its second security release in eleven days without saying what it fixes.
Component | Affected | Severity | Patched? | Action | Relevance |
|---|---|---|---|---|---|
mcp-atlassian ×18 | 9.6 · 8.1 · 7.7 · 7.2 · 7.1 · 1 high unscored · 9 medium · 3 low | yes → 0.24.1 (Nov 8) | upgrade; rotate the Atlassian token if HTTP/SSE was reachable beyond localhost | AI stack | |
Open WebUI | not published | yes → 0.13.1 (Nov 8); 0.13.3 current | upgrade to 0.13.3 | AI stack | |
ContextForge MCP Gateway | 6.5 | yes → 1.1.0 (Oct 20) | upgrade; the floor moves up from 1.0.8 | AI stack | |
aiohttp ×9 | 7.5 · 5 medium · 3 low | yes → 3.14.4 (Oct 5) | nothing new past 3.14.3; pin it where a dependency still allows less | AI stack |
A scanner will raise none of these today. The mcp-atlassian and aiohttp advisories are repository-only, Open WebUI has published no advisory, and GitHub's record for the ContextForge CVE names no package.
Worth your morning
mcp-atlassian: 18 advisories, two releases, one Sunday. mcp-atlassian (about 6,100 stars) connects agents to Jira and Confluence. It released 0.24.0 at 10:05 UTC and 0.24.1 at 15:23, and published 18 repository advisories with them: one critical, five high, nine medium, three low. All but one name 0.24.1 as the fix. Most matter only when the server runs over HTTP or SSE for several users, the mode the project documents for hosted, Kubernetes and multi-tenant setups.
- Confluence attachments (GHSA-cf68, 9.6): an unvalidated attachment ID let a caller of
download_attachmentread other Confluence REST paths with the service account's token, read-only mode or not, and in some cases send that token to another host. Over stdio, the caller can be a prompt-injected agent. - Project and space allowlists (GHSA-xv8h, 8.1):
JIRA_PROJECTS_FILTERandCONFLUENCE_SPACES_FILTERwere enforced on 2 of about 15 tools, so reads by key and every write reached projects outside the list. Where several users share one credential and the allowlist is the only boundary between them, that includes deletes. Thirteen duplicate reports of the same gap were closed into this one. - Files on the server (GHSA-qh7q, 7.7, CVE-2026-124503): the upload guard used the process's working directory as its boundary. An HTTP or SSE server started from
/, systemd's default, let a signed-in caller attach any file the process could read to a Jira issue or Confluence page, including the environment that holds the Atlassian tokens. - Requests without credentials (GHSA-rh5v, 7.2): in multi-user HTTP mode, the
X-Atlassian-Jira-UrlandX-Atlassian-Confluence-Urlheaders got a client without a token past the authentication gate and made the server send requests to hosts the client chose. Five more advisories (7.1 down to 4.9) close other routes around the SSRF guard: DNS rebinding through an operator's proxy, a request-URL header in external-auth mode, NAT64 addresses, the Forms API, and attachment downloads that forwarded credentials off-origin. - The rest: the OAuth setup helper (a callback listener on
0.0.0.0, a reflected XSS), secrets in debug logs, a JQL/CQL route around the allowlist, and two more path checks on downloads and issue links. - JSM internal notes (GHSA-q2h8, high): from 0.22.0 to 0.23.1,
jira_add_commentwithpublic=falseposted notes on Jira Service Management requests as customer-visible comments, and JSM emailed them to the customer. Deleting the comment doesn't recall the email. Fixed in 0.24.0; on older versions, listing the JSM projects inJIRA_INTERNAL_ONLY_PROJECTSkeeps notes internal. - The September floor: one medium (GHSA-2w5r) says 0.23.1, the release that fixed August's SSE 10.0, still left most of the SSE message endpoint ungated. Tool calls sent that way were refused, so operator credentials weren't used.
Update to 0.24.1. If the server ran over HTTP or SSE and was reachable beyond localhost, rotate the Atlassian API tokens or personal access tokens it was configured with. If agents comment on JSM projects through it, check their recent internal notes for any that went to customers. None of the 18 is in GitHub's database, and only GHSA-qh7q has a CVE, which hasn't reached NVD yet.
Open WebUI: another security release with nothing listed. 0.13.1 reached PyPI on Sunday at 11:15 UTC, 291 commits after 0.13.0. Like 0.12.5 on October 29, its notes say it carries security fixes that aren't in the changelog, ask everyone to update, and promise advisories once CVEs are assigned. The advisories for 0.12.5 haven't appeared either: the repository's newest is still from October 18. 0.13.2 and 0.13.3 followed the same day with bug fixes, so take 0.13.3. 0.13.1 also adds an "Auto" tool-permission mode in which a review model decides which tool calls run without confirmation; it ships off. And 0.13.3 still pins pypdf==6.21.2, below the 7.0.0 that fixes pypdf's five November advisories, and authlib==1.7.2, below the 1.9.1 that fixes Authlib's nine from October.
ContextForge: a new CVE moves the floor to 1.1.0. IBM's CVE-2026-18484 (6.5) reached NVD on Sunday. In ContextForge MCP Gateway 1.0.0 through 1.0.11, a user allowed to register gateways can register one whose tool names collide with an existing public tool in the same team, and calls to that tool then fail. IBM's bulletin (October 27) names 1.1.0, released October 20, as the fix; the floor this digest gave on October 14 was 1.0.8. Until you upgrade, renaming or removing one of the two colliding gateway registrations restores the tool, but nothing in the configuration prevents a new collision.
aiohttp: nine advisories for an October release. On Sunday evening aiohttp published nine advisories, all fixed in 3.14.4 (October 5). The highest (GHSA-3rp8, 7.5) lets a WebSocket peer get past max_msg_size in the C extension; the rest are memory and CPU exhaustion through form posts, multipart bodies, upgrade requests, compressed streams, wildcard host matching and the client cookie jar. Every advisory's patch commit is in the 3.14.4 tag, and 3.14.4 has been this digest's floor since October 7. vLLM 0.33.0, Ray 2.60.0 and the OpenAI SDK's aiohttp extra now require it. LiteLLM 1.108.3 still accepts 3.14.2 and langchain-community 0.4.2 anything from 3.8.3, so set the floor in your own lockfile.
Also on the AI stack.
- Conductor (conductor-oss, about 32,000 stars, a workflow engine with an AI agent runtime): VulDB's CVE-2026-124507 (5.3) reports SSRF in the agent planner's context-fetch task, up to 3.33.0 and 3.34.0-rc3. VulDB names no fix and the reporter's advisory isn't public. If you run Conductor's agent tasks, limit the server's outbound network access to what your workflows need.
- Chat2DB (about 28,000 stars, an SQL client that generates queries with your own model and supports MCP): VulDB's CVE-2026-124557 (7.3, 5.5 under v4) reports path traversal in the MyBatis class generator through 5.3.8. VulDB says the vendor didn't respond, and the file is unchanged in 5.3.9-beta.2. Keep the web and Docker editions away from untrusted users.
- IBM watsonx Orchestrate, the Cloud Pak for Data cartridge: four CVEs on Sunday, the highest 6.5, three of them backend APIs that skipped permission checks the UI enforced. Versions 4.8.4–4.8.5 and 5.0.0–5.4.2 are affected; IBM's bulletin (October 23) says upgrade to 5.4.4.
New paperwork, old fix. IBM's Langflow CVE-2026-97677 (8.1), one of the 24 from the October 2 bulletin covered here on October 7, reached NVD and GitHub's unreviewed feed on Sunday. It's fixed in 1.12.3, below the 1.13.0 floor that has applied since October 30.
On the Venicecom stack. MSRC revised nothing on .NET or Azure services since yesterday, and Patch Tuesday is tomorrow, 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 350 records were mostly VulnCheck batches for OpenEMR, Vikunja and OpenAM and Sunday's IBM bulletins (DataStage, Guardium, WebSphere, Qiskit); none touch the tracked stack beyond the items above.
Standing items. None of last week's repository advisories (GPUStack, OpenHuman, undici, Authlib, rmcp, Langflow, Jan, Better Auth, Next.js, Agent Router, pypdf, ToolHive, vLLM, DeepSeek-Reasonix, OpenCode) has reached GitHub's database. Reasonix 2.58.1 (overnight) brings no new advisories; its Studio build moves to Electron 45.2.3 for Chromium security fixes. Red Hat still lists the Kubeflow Trainer CVE as under investigation, and Trainer has no release after 2.4.0. LMDeploy's PR 4964, DeepInfer's PR 242, AgentScope's PR 3448, Dify's PR 44192, AG-UI's three client fixes and DB-GPT's issues 3287 and 3290 are still open. LlamaIndex's file readers and Unstructured, like Open WebUI, still hold pypdf below 7. FastMCP 4.1.6 still allows prefab-ui>=0.18.0, Gradio 6.31.0 still requires pillow<13.0, and 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. GitHub's MLflow record still names 3.17.2, and TorchServe's CVE still has no package mapping.
* * *
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.