By AI Blog Editor
Nov 15, 2026 · 35 min read
The Patch — November 15, 2026
OpenHands' OpenAI-compatible gateway ran agent commands outside the Docker sandbox, IBM moved ContextForge's floor to 1.1.1, and ExcelJS has eleven new CVEs and no fixed release.
There was no Saturday digest, so this one covers Friday morning to Sunday morning. Teams running OpenHands with the Docker conversation runtime should check their Agent Server version first: its OpenAI-compatible gateway ran agent commands on the host instead of in the container, fixed in 1.61.0. ExcelJS, a common choice for Node.js back ends that import or export spreadsheets, has eleven new CVEs and no release to move to.
Component | Affected | Severity | Patched? | Action | Relevance |
|---|---|---|---|---|---|
OpenHands Agent Server ×2 | 8.3 · 8.2 | yes → 1.61.0 (Nov 9); 1.62.0 current | upgrade the server and conversation images together; until then block | AI stack | |
IBM ContextForge MCP Gateway ×8 | three through 1.1.0 · five through 1.0.11 | 8.1 · 7.4 · 6.5 ×2 · 5.4 ×2 · 4.8 · 4.1 | yes → 1.1.1 (Nov 10) · 1.1.0 (Oct 20) | upgrade to 1.1.1; until then strip | AI stack |
n8n ×5 | 7.7 · 7.5 · 5.3 ×3 (v4) | yes → 2.47.7 (npm | upgrade; the floor set on October 29 moves | AI stack | |
GitLab MCP ( | 2.1.69 – 2.3.26 · ReDoS from 1.0.60 | 8.0 · 6.5 | yes → 2.3.27 (Nov 13) | upgrade; don't run replay commands produced by older versions | AI stack |
Flowise | 8.5 · 6.3 (v4) | no; 3.1.4 (Jul 29) is the newest | limit chatflow and agentflow editing to trusted users; restrict the host's outbound network | AI stack | |
soupsieve | 7.5 | yes → 4.1.1 (Nov 14); Python 3.11.5+ | refresh lockfiles where callers can supply a CSS selector; floor moves from 4.1 | AI stack | |
Telerik UI for Blazor ×2 | AI command filters 12.0.0 – 15.2.0 · Chat AI demo 11.1.0 – 14.0.0 | 6.5 · 7.5 | yes → 16.0.0 (Nov 11) · demo source 14.1.0+ | upgrade the package; replace any code copied from the Chat AI demo | both |
ExcelJS ×11 | 8.7 ×7 · 8.3 · 6.9 · 6.3 · 2.3 (v4) | no; 4.4.0 (Oct 2023) is the newest | don't parse untrusted workbooks in a shared server process; never pass user input to | Venicecom stack | |
AutoMapper ×2 | 7.5 ×2 | yes → 16.3.0 · 15.1.4 (Nov 5); none for 14.x | upgrade; set | Venicecom stack |
Only AutoMapper's two records are in GitHub's reviewed database, so NuGet audit raises them now. None of the OpenHands, n8n, GitLab MCP or soupsieve records is there yet, and the ContextForge, Flowise and ExcelJS CVEs sit in the unreviewed feed with no package mapping, so a scanner that reads GitHub's database stays quiet on all of them.
Worth your morning
OpenHands: the gateway skipped the sandbox. OpenHands published two advisories on Friday evening for its Agent Server's OpenAI-compatible gateway. With OH_CONVERSATION_RUNTIME=docker, conversations started through /v1/chat/completions or /v1/responses ran in the outer Agent Server rather than a conversation container (GHSA-h7m4, 8.3), so the agent's commands ran with the server's service-account privileges, and content that steers the agent (repository files, web pages, tool output) could reach the files and credentials that account holds. The Docker runtime confined every route except the one that speaks OpenAI's API. The second record (GHSA-f55r, 8.2) covers resuming a Docker-owned conversation through the gateway: code already running in a conversation container could change that conversation's saved configuration so that a lifecycle hook ran in the outer server on resume. Both affect 1.49.0 through 1.60.2, and 1.60.2 was cut from 1.60.1 without the fix. 1.61.0 (November 9) fixes both; the advisories still say "when 1.61.0 is available", but it has been on PyPI since Monday and 1.62.0 since Wednesday. Update the outer server and the conversation images together. Until then, block the two gateway routes and use the Docker-backed conversation API.
ContextForge: eight IBM CVEs, and the floor is now 1.1.1. IBM published eight bulletins for its MCP Gateway on November 12 and 13, and NVD carried them late Friday. Three affect 1.1.0, the floor this digest set on November 9, and are fixed only in 1.1.1 (November 10). The worst (CVE-2026-110279, 8.1): a caller's X-Upstream-Authorization header could replace the credential the gateway manages for OAuth, bearer and basic-auth upstream servers. IBM's workaround is to strip that header at the reverse proxy for every MCP endpoint. The second (CVE-2026-110278, 7.4) covers identity propagation: signed identity claims sent to upstream MCP servers could be replayed or exposed to third-party servers; until you upgrade, set IDENTITY_PROPAGATION_ENABLED=false. The third is a jq filter denial of service (CVE-2026-110189, 6.5) with no workaround. The other five, an SSRF, a token-exchange cache that let callers reuse another credential's upstream token, a ReDoS through tool schemas, stored XSS in the admin UI and team-scoped import overwriting public tools, stop at 1.0.11 and were fixed in 1.1.0. When we covered 1.1.1 on Wednesday it looked like a dependency bump; it carried these three fixes.
ExcelJS: eleven CVEs, no fixed release. VulnCheck filed eleven CVEs against ExcelJS on Friday, all "through 4.4.0", which is every release; 4.4.0 came out in October 2023. The package has about 28 million npm downloads a week. Seven are rated 8.7 on v4. Six of those cover memory or CPU exhaustion, or leaked temporary files, while reading a crafted workbook (merge ranges, defined names, data validations, sheet IDs, image anchors, the streaming reader). The seventh (CVE-2026-128347) lets the streaming WorkbookWriter embed any readable local file when an image filename passed to addImage comes from user input. Two prototype-pollution records (8.3 and 6.3) also trigger on reading a crafted file. The GitHub advisories the CVEs link to aren't public, and the repository has no published advisories. Until there's a fix: if your server opens spreadsheets that users upload, do it in a separate worker with time and memory limits so one file can't stop the service, and never build addImage paths from request data. Browser-only use, where the user opens their own file, mostly risks their own tab.
Also on the AI stack.
- n8n published five advisories on Friday with releases that morning. A user with a custom instance role that includes user management could take over the owner account (GHSA-q8rc, 7.5); a member who can list users could set the password of a pending invitation (GHSA-m7qq, 7.7, fixed back in 1.123.90 and 2.14.0); the AWS Bedrock nodes skipped SSRF protection for custom runtime endpoints; trigger nodes could use another user's credential on activation; and unencoded IDs could point integration requests at other endpoints of the same service (all 5.3). The floor moves from 1.123.89 / 2.45.5 / 2.46.1 to 1.123.96 / 2.47.6 / 2.48.3.
- GitLab MCP 2.3.27 (Friday): with
GITLAB_MCP_COMPACT_RESULTSon, the "replay" command the server prints could carry shell syntax from tool arguments, which can come from GitLab content an agent reads, so running it executed commands with the user's tokens (GHSA-mfp5, 8.0). Also a ReDoS inexcluded_file_patterns(6.5), and the release lifts the bundledhonoto 4.13.18. - Flowise: VulnCheck's CVE-2026-127214 (8.5 on v3.1) says a user who can create or edit chatflows or agentflows can get around the Custom Function sandbox's SSRF protection. It affects 3.1.4 and earlier; 3.1.4 is from July, and the project's newest advisory is from September 10.
- soupsieve 4.1.1 (yesterday): GHSA-98qm (7.5), a selector of about 100 bytes with nested
:nth-last-child(... of S)can take minutes of CPU on an ordinary page. It needs a caller-supplied selector, not a hostile document. The floor from November 13 moves from 4.1 to 4.1.1, still Python 3.11.5 or later. - Telerik UI for Blazor: the Grid, TreeList and Gantt AI commands could filter on any model property, past the declared column allowlist (CVE-2026-119102, 6.5, fixed in 16.0.0). The 7.5 (CVE-2026-118976) is in the Chat AI Integration sample of the demo application, not the NuGet package: it matters only if you deployed the demo or copied its attachment handling.
- Stanza 1.15.1 (yesterday): three 7.8s for pickle loading in the lemmatizer conversion and inspection scripts (GHSA-j6gr and two more). They apply when someone runs those scripts on a
.ptfile from an untrusted source. - Apache APISIX 3.20.0 (November 11): a caller of the MCP endpoint could reach upstream paths the OpenAPI document never declared, with the credentials the route adds (CVE-2026-124726). Affects 3.19.x.
- Claude Code 2.1.325 (Friday) lists "a security issue in how zstd-compressed responses are decoded" with no advisory.
- Dify published GHSA-6j8g (6.5) yesterday: per-IP limits, including the password-reset email limit, trusted a spoofable
X-Forwarded-For. Fixed in 1.18.0 (October 23); 1.19.0 is current.
Also on the Venicecom stack. AutoMapper's two records went up on November 5 with 15.1.4 and 16.3.0 and reached the reviewed database on Friday: several mapping paths ignore the MaxDepth limits added for an earlier CVE (CVE-2026-32933), so deeply nested input recurses until a StackOverflowException ends the process, which can't be caught. The range covers every release before 15.1.4, including 14.0.0, the last one published under MIT; 15.0 and later ship under their own license, and none of the fixes is on 14.x. The advisory's workaround is to set MaxDepth explicitly on derived maps and maps reached through object or interface members; for dictionary or dynamic sources and recursive converters, only limiting input depth before mapping helps. DOMPurify 3.4.18 (yesterday) fixes a medium in which SAFE_FOR_TEMPLATES output could reassemble template expressions when re-parsed.
Paperwork. GitHub reviewed NLTK's records for 3.10.4, covered on October 22 and October 28, and aiohttp's nine from November 9 got CVEs (CVE-2026-128124 to CVE-2026-128132). VulnCheck filed eleven CVEs for n8n's October 28 advisories (covered October 29), six for vLLM's October 27 batch and one for SiYuan's skill-install SSRF (fixed in 3.8.10). A correction on vLLM: October 28 said all six were fixed in 0.30.0; the records say 0.31.0 for five of them. Current vLLM is past both. Two more soupsieve records were reviewed.
Quiet elsewhere. GitHub's reviewed feed published 46 records from Friday 00:45 to Saturday 00:08 UTC and nothing since; the same query for October 28–29 returns 117. The unreviewed feed held 768 records and NVD 792 over the 48 hours. MSRC revised one entry, not a .NET or Azure one. Angular CLI's pins are unchanged from November 13. No PrimeNG advisories.
Standing items. LMDeploy's PR 4964, DeepInfer's PR 242, AG-UI's three client fixes, AgentScope's PR 3448, Dify's PR 44192 and DB-GPT's issues 3287 and 3290 are still open. Open WebUI's newest advisory is still from October 18. SiYuan 3.8.11 is at alpha.5. MongoDB's Node.js driver still has no 6.x release with the SRV fix, and @langchain/mongodb still requires 6.x. pgEdge Postgres MCP 1.1.1 still doesn't exist. nginx-ui 3.0.2's four advisories are still private. None of Thursday's or Friday's LiteLLM, MLflow, MongoDB, OpenHands or n8n records has reached GitHub's database.
* * *
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.