News · Security · November 19, 2026
LiteLLM: a fourth CISA-listed flaw in 2026. What to do
CISA listed LiteLLM CVE-2026-40217 as exploited on 18 Nov, the fourth this year. Patch to 1.83.11 or later, then rotate the keys the proxy holds.
DRIVE 369 DAO · 4 min read
On 18 November CISA added a fourth LiteLLM vulnerability of 2026 to its Known Exploited Vulnerabilities catalog: CVE-2026-40217, a sandbox escape that lets code run inside the proxy. LiteLLM is an open-source gateway that teams put in front of their AI providers, so it usually holds their provider keys. US federal agencies must act by 2 December. Everyone else should upgrade to 1.83.11 or later, then rotate the keys the proxy holds if it was reachable.
What is stated
| Question | Answer | Source and tier |
|---|---|---|
| What is the flaw? | The endpoint POST /guardrails/test_custom_code runs user-supplied Python in a hand-rolled sandbox. Bytecode-level techniques escape it and run arbitrary code in the proxy process, which runs as root in the default Docker image | LiteLLM advisory GHSA-wxxx-gvqv-xp7p, vendor, reported |
| Who can reach it? | Reaching the endpoint needs a proxy-admin credential in default configurations. The advisory rates the severity medium | Same advisory, vendor, reported |
| Which versions? | 1.81.8 up to, but not including, 1.83.10 | Same advisory |
| Which fix? | The advisory text says 1.83.11 and the sandbox was replaced with RestrictedPython; its machine-readable field says 1.83.10 or later. The two disagree by one release | Same advisory; we advise 1.83.11 or later, which satisfies both |
| Is it exploited? | CISA lists it as known exploited, added 18 November 2026, due 2 December | CISA KEV catalog, primary |
| When was it public? | The advisory was published on 7 May 2026 and updated on 30 June, about six months before the CISA listing | Advisory metadata, primary |
CISA's entry calls it an unprotected alternate channel (CWE-420) and says remote attackers could execute arbitrary code. The advisory adds the admin-credential condition. Both can be true: the listing says the flaw is used in the wild, and the advisory says what an attacker needs first.
Four listings, four different flaws
The catalog counts vulnerabilities with evidence of exploitation, not incidents. These are four separate flaws in one product:
| Added | CVE | What CISA says |
|---|---|---|
| 8 May 2026 | CVE-2026-42208 | SQL injection: an attacker can read data from the proxy's database and potentially modify it |
| 8 June 2026 | CVE-2026-42271 | Command injection: any authenticated user, including holders of low-privilege internal-user keys, can run commands on the host |
| 2 September 2026 | CVE-2026-59822 | Improper authentication in the MCP Streamable HTTP endpoint |
| 18 November 2026 | CVE-2026-40217 | Code execution through the guardrails test endpoint |
Source: CISA's catalog data, release of 18 November 2026 (1,786 entries). Do not read it as one attacker or one campaign: the entries do not say.
Why a gateway is a high-value target
A gateway sits between your applications and every AI provider you pay. It holds the provider keys, the keys it issues to your teams, and often the logs of what was asked. Code running as root inside it can read all of that. That is why where AI keys should live matters more for the gateway than for any single app.
Do this week
- Find every LiteLLM proxy you run, including staging and the one a colleague set up for a demo. Check the installed version or the image tag.
- Upgrade to 1.83.11 or later. If you cannot yet, block POST /guardrails/test_custom_code at your reverse proxy or API gateway, which is the maintainers' workaround.
- Rotate after you patch, not before. Rotating on a vulnerable proxy hands the new keys to the same hole. Then rotate the provider keys stored in it, the proxy's master key and every admin credential, if the proxy was reachable from the internet at a vulnerable version.
- Shrink who is admin. The flaw needs an admin credential, so every shared or old one is a way in. Keep the proxy behind a VPN or single sign-on, not on the open internet.
- Look for signs of use. Check the proxy's logs for calls to the endpoint and your providers' dashboards for spend you did not make. Amazon's advice in its ARTEX report is similar: no default credentials, the same controls for AI gateways, and a watch on unexpected spend.
Where Ghostdrive stands
Ghostdrive does not run a gateway and cannot patch yours. If your provider keys live in a LiteLLM proxy, they are as safe as that host. Pass keeps API secrets as sealed items on your device, and a scoped key limits the damage of one leak. Neither fixes a vulnerable proxy. Checks for the habits around keys are in ten guardrails every AI agent needs.
Limits
We have not reproduced the flaw. CISA's entry gives no detail on who exploited it, how many were affected or when exploitation was first seen, and we do not know whether the four flaws were used by the same actors. The advisory's severity and CISA's listing weigh the same facts differently; we show both.
Academy: where AI keys should live · Before you connect an MCP server · Black October explained
Questions
- Do I need a login to be hit by CVE-2026-40217?
- The maintainers say reaching the endpoint needs a proxy-admin credential in default configurations. CISA lists the flaw as exploited but does not say how. Treat an exposed proxy with shared or old admin keys as at risk.
- Which version fixes it?
- The advisory text says 1.83.11, while its machine-readable field says 1.83.10 or later. Upgrading to 1.83.11 or later satisfies both. Affected range: 1.81.8 up to, but not including, 1.83.10.
- Does CISA say who is exploiting it?
- No. The catalog entry gives the flaw, the required action and the dates, and nothing about the attackers, the victims or when exploitation was seen.