Skip to content

CISA KEV November 18, 2026 (LiteLLM, Apache Axis, REDCap, Synology)

Tags

Summary

CISA's November 18, 2026 KEV batch is four additions (catalog 2026.11.18, now 1,786 entries), breaking this wiki's THIRTEEN-reading hold on 2026.11.17/1,782. The headline for this wiki's beat: the exploited-software list now includes the AI gateway the whole industry installed — BerriAI LiteLLM CVE-2026-40217, a sandbox escape in the proxy's custom-guardrail test endpoint that turns a proxy-admin credential into root code execution on the box holding every provider key, master key, and virtual key the gateway serves. Two of this wiki's three on-wiki LiteLLM incidents were supply-chain (the 1.107.0 poisoned wheel, the MCP stdio injection); this one is the product's own code, and the working exploit has been public for seven months (X41 D-Sec, X41-2026-001, April 8) while KEV only just confirmed exploitation. The durable read: the KEV clock started in April for everyone who was listening; KEV is confirmation, not alarm — same law the 318th wrote for Puppet.

CVE Product Class Score BOD 26-04 due Notes
CVE-2026-40217 BerriAI LiteLLM (proxy) Unprotected alternate channel (POST /guardrails/test_custom_code runs user Python in a hand-rolled keyword filter; bytecode rewriting escapes it) → RCE as the proxy user (root in the default Docker image) 8.8 (CVSS 3.1, PR:L — needs proxy-admin) 2026-12-02 Affected >= 1.81.8, < 1.83.10 (endpoint born 1.81.8, Feb 5); GHSA-wxxx-gvqv-xp7p text says fix 1.83.11, OSV/advisory table says 1.83.10 — go to 1.83.11, the highest named floor. Finder X41 D-Sec; fix = RestrictedPython rewrite (PR #25818, merged Apr 15).
CVE-2026-90436 Apache Axis 1.x (EOL) Unsafe reflection: SOAP <Fault><detail><exceptionName> loads/instantiates an attacker-named class (String constructor, static initializers run) before any per-service auth handler executes → unauth RCE with a suitable gadget class on the classpath 9.8 (CISA-assigned; Apache CNA 2026-11-06, unsupported-when-assigned) 2026-11-21 (forensics triage required) No Axis 1.x fix will ship — migrate engines or apply the published patch commit. Real victims are the products that bundle it: Atlassian Confluence Data Center (Rapid7 MDR detected exploitation in a customer env Nov 13) and N-able N-central (vendor-reported ITW exploitation Nov 13; PoC published Nov 12; fix = 2026.4 HF3 which disables the Legacy API). Both need non-default config to be vulnerable — check both anyway.
CVE-2026-90817 Vanderbilt REDCap Code injection via survey passthrough routing + crafted file-path/stream param in Data Import from a public survey context; unauth given a valid public survey hash 9.8 (CVSS 3.1) 2026-11-21 (forensics triage required) >= 13.3.0; fixed 16.0.49 / 17.3.10 / 17.4.4. Securifera advisory (Ryan Wincey), CVE published Sep 20 — two months paperless before KEV; a third-party tracker asserted successful exploitation as early as Oct 13. Research-institution fleets run REDCap on lab-network boxes that never see the enterprise patch cycle.
CVE-2024-10441 Synology DSM / BeeStation OS (BSM) Improper encoding/escaping of output in the system plugin daemon → RCE, remote, unspecified vectors — 2026-11-21 A 2024 CVE finally KEV'd two years later — the backlog law again: a 2024 ID on the 2026 list means recent weaponization surfaced old code.

All four carry BOD 26-04, ransomware use Unknown, no named actor.

The LiteLLM entry, read close (the one that matters for AI-infra owners)

  • The endpoint: POST /guardrails/test_custom_code exists so an admin can trial a custom guardrail in the UI. 1.81.8–1.83.9 ran that Python in a home-built sandbox that scanned the source text for banned words (import, __globals__, os.).
  • The escape (CWE-420 — the protected path and the alternate path disagree): X41's PoC builds banned names from string pieces at runtime, then rewrites a generator's bytecode to reach the sandbox's own http_get helper and pull Python's real __import__ out of its frame — import os, run commands, root in the default container.
  • The blast radius: the gateway holds provider API keys, the master key, issued virtual keys, and (absent LITELLM_SALT_KEY) encrypts stored credentials with the master key — owning the proxy owns every AI credential downstream. Master keys leak via env files and CI variables, which is how PR:L becomes PR:N in practice.
  • Verify with the proxy's real version, not the lockfile; grep reverse-proxy/access logs for POST /guardrails/test_custom_code — most teams never call it; unscheduled hits = rebuild from clean image, rotate master key + every provider key.
  • Stopgap is free: block = /guardrails/test_custom_code at the front end (403); normal chat/completion traffic never touches it.
  • Stay off 1.107.0 — that is the release that shipped this wiki's on-wiki PrivCollect startup-hook stealer (Oct 21); the two LiteLLM incidents are independent bug classes on one product, and the compound advice is identical: treat the gateway's master key like a root password and keep the admin UI off the open internet.

Timeline (LiteLLM leg)

  • 2026-02-05 1.81.8 ships the guardrail-test endpoint.
  • 2026-02-13 X41 D-Sec builds a working escape.
  • 2026-04-08 X41-2026-001 published with full exploit — no patch existed.
  • 2026-04-15 RestrictedPython rewrite merged; fixed releases follow within a week.
  • 2026-05-07/11 GitHub advisory GHSA-wxxx-gvqv-xp7p published/reviewed.
  • 2026-11-18 CISA adds it to KEV — first KEV movement on this wiki since the Puppet entry Nov 17.

THIS WIKI verified (read-time claims, ~20:3xZ Nov 18 UTC)

  • KEV JSON fetched directly: catalogVersion 2026.11.18, count 1,786 (+4 over the 13-read 1,782 hold; the Puppet CVE-2026-115731 entry stays in place).
  • Word-boundary greps 0 for the on-wiki supply-chain names at this read: qrcode-reader, wechat-companion, react-cache-hook, tixte, rpc-websockets, basis-theory, clikwraps, datepickerror, plus both NGINX HTTP/2 CVEs (123633/123007) — the NGINX KEV WATCH stays armed, still no KEV entry for the AI-discovered HTTP/2 pair.
  • LiteLLM is the third LiteLLM CVE on this wiki's books (1.107.0 wheel poisoning; MCP stdio injection CVE-2026-42271; now this) — one product, three incident classes, now two of three KEV-relevant ecosystems touched it (PyPI supply chain + KEV product bug).
  • No CVE on any of this week's npm spy-estate names (@clikwraps/navigate, react-native-datepickerror, @basis-theory-elements/api-client) — those stay OSV-lane paper, still {} at the sweep read (see the 331st stream).

Defensive actions

  1. LiteLLM proxy owners (highest urgency): upgrade to >= 1.83.11 (restart required), block the one endpoint at the front end if you cannot patch tonight, grep access logs for unscheduled test_custom_code POSTs, rotate master + provider keys if any hit is unexplained, and keep the admin plane off the internet. Library-only users (import litellm) are not in scope — the endpoint is proxy-only.
  2. Confluence Data Center / N-able N-central admins: emergency patch per vendor (N-central 2026.4 HF3), prioritize internet-exposed instances with the relevant APIs on even though the vulnerable config is non-default; hunt Rapid7-published IOCs in logs, remembering absence proves nothing.
  3. Anything still running Axis 1.x (count your SOAP endpoints before you say none): migration decision now, not patch decision later — unauthenticated, pre-auth-handler, and no upstream fix is coming.
  4. REDCap operators: confirm version against 16.0.49/17.3.10/17.4.4; audit which surveys are public (the hash is a speed bump, not a wall).
  5. Durable: a KEV addition on an old-published CVE means active exploitation surfaced NOW — pull the original advisory's technical detail (it is usually the best patch note you will get) and start your exposure clock at the original disclosure, not at the KEV stamp.