On Wednesday, November 18, CISA added LiteLLM flaw CVE-2026-40217 to its Known Exploited Vulnerabilities catalog, citing evidence that attackers are already using it. Google's threat intelligence team reported that the bug lets attackers take over AI gateway servers, the proxies that hold the API keys connecting a company to OpenAI, Anthropic and Google Gemini accounts. Federal agencies now have until December 2, 2026 to fix it.

The bug is months old. It was disclosed in April 2026 and the fix was merged on April 15, so many self-hosted LiteLLM deployments may still be running vulnerable code. A flaw with a fix available since spring is now confirmed as an active target, and that gap is what makes the listing urgent.

If you run LiteLLM, the core steps are short. Check your version, upgrade to 1.83.14-stable, and block the vulnerable endpoint if you cannot patch right away. The rest of this article explains why, and where the sources still disagree.

Why a KEV listing matters for AI gateway operators

LiteLLM is an open-source proxy that gives developers one API across more than a hundred model providers. According to NFLO Tech, it also handles load balancing, cost tracking and guardrails. Many teams place it in front of several model accounts so that applications never hold provider keys directly.

That design concentrates risk. Because the proxy stores every upstream key, one compromised gateway can expose every connected model account at once, as Beri and Nexgismo both point out.

CISA's catalog is narrower than a general vulnerability database. It lists flaws with evidence of active exploitation, and its entries carry federal fix deadlines. A CISO briefing from Cloud Security Alliance Labs counts this as LiteLLM's fourth entry in the catalog in 2026. The same batch included three other flaws: Synology DSM CVE-2024-10441, Apache Axis CVE-2026-90436 and Vanderbilt REDCap CVE-2026-90817, according to SC World. USA 24 has already covered the Synology flaw in a story on C3RB3R ransomware hitting NAS devices running outdated DSM.

How attackers turned a limited key into admin access

Google's threat intelligence team reported that a threat actor used CVE-2026-40217 to convert limited-access inference keys, which LiteLLM calls Virtual Keys, into full administrative credentials. That activity falls under LLMjacking, the practice of hijacking AI model access for resale or for the attacker's own use. The Hacker News reported that the batch of four flaws is linked to ransomware, cryptomining and LLMjacking attacks.

USA 24 could not retrieve Google's original write-up. The Google details in this story come from secondary reporting, including a DEV Community analysis, and should be checked against the original before they are repeated in formal reporting.

The flaw itself is a sandbox escape in the custom-code guardrail test endpoint, POST /guardrails/test_custom_code. A proxy admin can run code on the server through that endpoint. Ostorlab's analysis notes that the proxy process runs as root in the default Docker image, and Decryption Digest scores the flaw at CVSS 8.8.

Zenity Labs observed attack traffic hitting that exact endpoint. The payloads rebuilt blocked keywords at runtime, a way to slip past the sandbox's filters, and then tried to read secrets and run commands. That detail gives defenders a specific request pattern to search for in their logs.

The attack path is the part many teams will miss. The proxy holds keys for many services, so a single foothold in its admin functions reaches all of them. Treat any unexpected admin or key-regeneration activity as a possible incident, even before you know which chain an attacker used.

Which LiteLLM versions are affected and which release to install

The affected range runs from LiteLLM 1.81.8 up to the first fixed release, but sources disagree on where that cutoff sits. Zenity lists the range as up to but not including 1.83.10. Decryption Digest says before 1.83.11, and Nexgismo gives 1.81.8 through 1.83.10. Decryption Digest reports that 1.83.11 shipped on November 15, 2026.

Item What the sources say
KEV addition November 18, 2026
Federal fix deadline December 2, 2026
Disclosure April 2026; Zenity dates it April 10
Fix merged April 15, 2026
Severity CVSS 8.8
Fixed release 1.83.10 (Zenity), 1.83.11 (Decryption Digest)
Recommended upgrade 1.83.14-stable (Beri)
Companion bugs CVE-2026-47101 and CVE-2026-47102

Beri describes 1.83.14-stable as the first complete fix, covering the companion privilege-escalation bugs. Those two flaws are described as part of the same chain. For most teams, the practical move is to skip the minimum fixed release and go straight to 1.83.14-stable.

How to check and patch your LiteLLM proxy this week

Work through these steps in order. Each one reduces exposure even if the others are not finished.

  1. Check the installed version. Look at your pip environment, your Docker image tag or your Helm chart values. Anything from 1.81.8 up to the version before the fix falls in scope.
  2. Upgrade to 1.83.14-stable. Do not stop at the minimum fixed release, because sources disagree about it and Beri's guidance covers the companion bugs.
  3. If you cannot patch now, block POST /guardrails/test_custom_code at a reverse proxy or API gateway, and restrict API access to trusted networks, as NFLO Tech recommends.
  4. Review admin accounts and guardrail configurations for changes you did not make. Search logs for requests to that endpoint.
  5. If you find misuse, rotate every provider key and credential the gateway held, not just the ones you think were touched.

Before you install any LiteLLM package, check the official PyPI and GitHub release pages. A CSA briefing claims a poisoned litellm==1.107.0 release is circulating, but USA 24 could not confirm that claim in any other source.

Exposed keys are a recurring problem in developer tooling. USA 24 has covered a Claude Code configuration file that exposed secrets in 1,122 public GitHub copies. Keep gateway configs and key files out of shared repositories, and rotate any key that has appeared in one.

Open questions and conflicting details

  • Attribution: No threat actor or campaign has been named publicly in the sources reviewed. The scale of exploitation and the number of victims are unknown.
  • US impact: No US-specific victim count or breach notification has been found. The direct exposure is to businesses and developers that run self-hosted gateways holding provider keys, plus federal agencies facing the December 2 deadline.
  • Mechanism: The CISA listing and several write-ups describe a sandbox escape. Beri describes a separate chain in which a user-ID rebind in the key-regeneration endpoint lets a standard virtual key become a proxy admin. The two may be parts of one chain, and the sources do not settle it.
  • Dates: Beri dates the Google and Mandiant report to November 2, while the DEV Community article dates Google's write-up to November 17.
  • Fixed release: Sources name 1.83.10, 1.83.11 and 1.83.14-stable as the fix. The newest of these is the most conservative choice.
  • Unverified release: A CSA briefing says a poisoned litellm==1.107.0 release is circulating. USA 24 could not verify it elsewhere.

What to watch before the December 2 deadline

Three developments will show how serious this campaign is. The first is whether CISA, Google or another researcher names an actor or publishes a victim count, which the current sources do not provide. The second is whether the companion bugs CVE-2026-47101 and CVE-2026-47102 receive their own catalog entries or detailed reporting, which would confirm that the chain is larger than one flaw. The third is whether the poisoned-release claim is confirmed or dismissed, since that would change how teams should pull LiteLLM packages.

The federal clock is a familiar pattern. Brotli2Shell, which was also being exploited, carried a November 10 federal deadline. Private companies face no such rule, but attackers do not follow the government calendar either. Teams that run AI gateways should treat December 2 as their own deadline.