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 16, 2026 · 24 min read

The Patch — November 16, 2026

NLTK published twenty more advisories for its October release, one of them a 9.8, and mcp-atlassian 0.24.5 fixes six findings from an internal audit.

NLTK published twenty advisories on Sunday, one rated 9.8, and every one is fixed in 3.10.4, the October 21 release this digest has carried as the floor since October 22. If you held at 3.10.3 because 3.10.4 changed loading behaviour, that hold now leaves a 9.8 open. Teams running mcp-atlassian over HTTP, or relying on its project and space filters, should take 0.24.5 from this morning.

Component

Affected

Severity

Patched?

Action

Relevance

NLTK ×20

≤ 3.10.3; the 9.8 on 3.10.3 only

9.8 · 7.5 ×3 · 7.3 · 7.1 · 1 high unscored · 13 medium

yes → 3.10.4 (Oct 21)

upgrade anything still below 3.10.4; the floor doesn't move

AI stack

mcp-atlassian ×6

< 0.24.5

1 high · 5 medium (no CVSS)

yes → 0.24.5 (Nov 16)

upgrade; until then set MCP_ALLOWED_URL_DOMAINS and use a credential scoped to the intended projects

AI stack

VoAPI

1.0.0 – 1.2.0

8.8 · 8.7 (v4)

no; 1.2.0 (Jan 20) is the newest

give accounts only to people you would trust with the management API

AI stack

None of the 26 NLTK and mcp-atlassian records is in GitHub's reviewed database, and NLTK's CVE IDs are assigned but not yet published at cve.org, so NVD has nothing either. For NLTK the gap costs little: an October record already in the database flags everything up to 3.10.3. For mcp-atlassian a scanner stays quiet below 0.24.5. The VoAPI CVE sits in the unreviewed feed with no package mapping.

Worth your morning

NLTK: twenty more advisories for a release from October. NLTK's maintainers published twenty advisories between 12:24 and 12:36 UTC on Sunday, each with a CVE ID (CVE-2026-128135 to CVE-2026-128140 and CVE-2026-129221 to CVE-2026-129234). All twenty name 3.10.4 as the fix. That release went out on October 21 and now closes 33 advisories: two covered on October 22, eleven on October 27 and these.

The 9.8 (GHSA-x8q7) affects 3.10.3 only. That release tightened how NLTK launches Java tools, and the fix was incomplete: where an application lets untrusted input reach the arguments NLTK passes to a Java tool, that input could run Java code with the application's privileges. The score assumes the input gets there. The report's own text suggested "moderate", depending on how often applications expose those arguments.

The six highs split in two. Four stall or crash a worker on crafted input: an XML corpus file, a dependency treebank, a custom WordNet resource, or a caller-supplied regular expression on 3.10.0 to 3.10.3 (GHSA-j8g8, GHSA-rh8m, GHSA-8846, GHSA-32p6). Two need a local user on the same host: package extraction could overwrite a file outside NLTK's data directory (GHSA-wr3g, 7.1, from 3.9.4), and a prover9 or mace4 binary configured with a relative path could resolve to a file in a directory that user can write to (GHSA-cc5r, 7.3). The thirteen mediums are mostly the same two patterns, recursion or quadratic time on crafted corpus files, plus a file read or create when an untrusted model goes through the restricted pickle loader (6.1).

Several advisory texts still say no fixed version exists; they were written in September and published in November. The structured field says 3.10.4, and the fix commits are in the 3.10.4 tag. The exposure is a service that runs NLTK over text or corpus files it didn't produce, which includes ingestion pipelines: llama-index-core requires nltk>=3.9.3 with no upper bound, so refreshing the lockfile is enough. As October 22 noted, 3.10.4 restricts loading to the directories in nltk.data.path and raises "Security Violation" errors where code relied on the old behaviour.

mcp-atlassian 0.24.5: six findings from an internal audit. The MCP server for Jira and Confluence (about 625,000 PyPI downloads a week) released 0.24.5 at 04:18 UTC today and published six advisories three minutes later. None carries a CVSS score; the labels are the maintainer's. The advisories say they come from an internal audit of the project. It is the second batch in eight days, after the eighteen fixed in 0.24.1 and covered on November 9.

  • The high (GHSA-f5qp): with JIRA_PROJECTS_FILTER or CONFLUENCE_SPACES_FILTER set, ID arguments on several tools could address objects outside the allowlist. A caller, or a model steered by content it read, could read them and, with write tools enabled, change them. It stays within what the server's credential can do. READ_ONLY_MODE=true blocks the write half.
  • GHSA-vhwv: the same filters were skipped for empty arguments, and changelogs, comment bodies, templates and space lookups could return content from outside them.
  • GHSA-qx98: with ALLOW_GLOBAL_CRED_FALLBACK=true or IGNORE_HEADER_AUTH=true on an HTTP transport, the server didn't check Host and Origin, so a web page in a browser that can reach the listener could get requests answered with the operator's credentials. Until you upgrade, set FASTMCP_HTTP_HOST_ORIGIN_PROTECTION=true or turn both settings off.
  • GHSA-cvhg: on HTTP transports where callers send their own URL and token, attachment downloads weren't tied to the Jira instance of the request, and JIRA_SSL_VERIFY=false also switched off certificate checks for hosts the caller named.
  • Two availability issues: one slow upstream call could stall a shared HTTP server for every caller, and crafted Jira or Confluence markup could do the same.

A single-user stdio install is largely outside the four HTTP findings. If you use the filters as a boundary, the first two apply on any transport, and the durable fix is the advisory's own workaround: a credential whose Atlassian permissions match the intended scope. The floor moves from 0.24.2 (November 12) to 0.24.5.

VoAPI: an 8.8 with no fixed release. VoAPI is a gateway that aggregates LLM provider APIs and hands out access through its own accounts and keys (about 1,100 stars). VulnCheck's CVE-2026-129540 (8.8; 8.7 on v4) says any signed-in user can take over the management API. It covers 1.0.0 through 1.2.0. The newest release is 1.2.0 from January 20, the repository has had no push since January 27, and the public report was opened and closed on Sunday within four minutes, with no release since. Until one appears, an account on a VoAPI instance is an administrator account in practice: limit sign-up to people you trust, and rotate the upstream provider keys if that wasn't the case.

Also on the AI stack.

  • AutoGPT Platform 0.11.3 (Sunday): the notes say it "fixes several privately reported security issues" and that advisories will follow once the fixes are live in production. Self-hosters should upgrade without waiting for them. The floor from November 13 moves from 0.11.2 to 0.11.3.
  • LightRAG 1.5.8rc1 (Sunday), no advisory: through 1.5.7, the default WHITELIST_PATHS is /health,/api/*, so the Ollama-compatible /api/* routes answer without the key even when LIGHTRAG_API_KEY is set. The release candidate requires the key on them. Stable is still 1.5.7 (September 2), so set WHITELIST_PATHS=/health there. The candidate also moves PDF extraction into a resource-limited child process and needs Python 3.11 or later.
  • OpenTelemetry Python API 1.45.2 (October 29): GHSA-mf5v (CVE-2026-126360, 5.3), reviewed on Friday and not covered here: a crafted traceparent header takes quadratic time to parse, a denial-of-service risk for any service that accepts trace headers from outside. It's in GitHub's database, so pip-audit raises it. AutoGPT 0.11.3 and code-graph-rag 0.6.62 both bumped for it over the weekend.
  • vLLM: VulDB filed CVE-2026-129356 (5.0; 1.3 on v4) for a race condition in KV cache prefix handling on 0.30.0 and 0.31.0. The record names 0.32.0 as the fix, the pull request merged on October 12, and 0.33.0 is current.

Quiet elsewhere. Nothing landed for Angular, PrimeNG, .NET or Azure. GitHub's reviewed feed has published nothing since Saturday 00:08 UTC; the same query for October 28–29 returns 117, so the zero is the weekend. The unreviewed feed held 107 records and NVD 113 over the 25 hours, mostly WordPress plugins, D-Link and DrayTek routers, Assimp and GPAC. MSRC revised two libslirp entries and nothing for .NET or Azure. Angular CLI's pins are unchanged from November 13.

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.6. ExcelJS and Flowise still have no fixed release. 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 last week's OpenHands, n8n, GitLab MCP, soupsieve, Stanza, Dify, LiteLLM, MLflow or MongoDB 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 more
  1. 01

    News

    Anthropic's day one as ANT — opened $111.50, closed $97.17, founders cashed $1.96 billion

    Nov 20, 2026

  2. 02

    The Patch

    The Patch — November 20, 2026

    Nov 20, 2026

  3. 03

    News

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

    Nov 19, 2026

Letters

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