CISA Lists AI Gateway LiteLLM Vulnerability CVE-2026-40217
CVE-2026-40217, a sandbox escape in the LiteLLM proxy, was added to CISA's KEV catalog this Wednesday with a federal patch deadline of December 2.

On Wednesday (18), CISA added four vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog, the list of issues that the U.S. agency considers actively exploited. One of them stands out for those managing AI in production: CVE-2026-40217, in LiteLLM, the open-source proxy from BerriAI that many companies use as a single gateway to access models from OpenAI, Anthropic, Google, and other open weight providers. The agency classified the flaw as "Unprotected Alternate Channel," and federal civilian agencies have until December 2 to correct it.
The other three entries in this batch have a much shorter deadline of November 21: CVE-2026-90436, in Apache Axis, an outdated SOAP library still present in legacy systems; CVE-2026-90817, in REDCap, a research database from Vanderbilt University used by hospitals and universities; and CVE-2024-10441, in Synology's DiskStation Manager, corrected in 2024, which the manufacturer recently warned was being exploited to install ransomware on outdated NAS devices.
How the Flaw Works
According to the security advisory from the project and analyses published by researchers, the problem lies in the code path used to test custom guardrails, the rules that filter what enters and exits the models. An attacker can escape the sandbox of this test and execute their own code on the server; in the standard Docker image, this code runs as root. Exploitation, in default configurations, requires an administrator key from the proxy, which reduces the attack surface but does not eliminate it: gateway keys often circulate in CI pipelines, configuration files, and internal repositories.
The Google Threat Intelligence Group (GTIG) recently reported on an actor who utilized exactly this vulnerability to turn limited access keys into administrator credentials. It is this real use that sets CVE-2026-40217 apart from merely theoretical vulnerabilities and justifies its inclusion in the KEV.
Sources differ on the severity: one database lists a CVSS score of 8.8, while another records 5.9 and notes that the LiteLLM advisory uses version 4 of the metric. The difference reflects the weight given to the requirement for the administrator key. For those deciding correction priority, the score matters less than the fact that exploitation is confirmed by CISA.
Why an AI Gateway is a Preferred Target
A proxy like LiteLLM centralizes all API keys from the company’s model providers, prompt and response logs, and, in many cases, the guardrail rules approved by the risk department. Whoever controls this server controls spending on tokens, reads what employees send to the models, and can disable filters without leaving an obvious trace in the application. The GTIG report on AI agents automating exploitation and credential theft specifically highlights this type of component as a new, poorly monitored link.
The component typically enters through the engineering door, installed via pip or Docker image by a platform team, and not through the procurement process. This means many CISOs are unaware of how many instances exist, who holds the administrator keys, or whether they are exposed to the internet.
The Effect on Multiple Markets
In the United States, the December 2 deadline formally applies only to federal civilian agencies under directive BOD 22-01, but government suppliers and regulated banks use the catalog as a prioritization measure, and auditors are starting to inquire about it.
In the European Union, companies subject to NIS2 and banks under DORA are required to manage vulnerabilities in third-party components, including open source. A compromised AI gateway that leaks prompts with personal data also triggers reporting obligations under GDPR.
In India and Brazil, where integrators and shared service centers build and operate a considerable amount of generative AI applications for global clients, correction depends on who controls the environment. In managed service contracts, the practical question for the client is whether the provider can identify, by the deadline, in how many projects LiteLLM is installed and who holds the administrator key.
The project’s recommendation is to update to the patched version indicated in the advisory. Where this is not immediate, analysts suggest blocking the guardrail testing endpoint in the proxy itself and changing the administrator key. Changing the key is often something that gets deferred, but it is the measure that closes the path outlined by GTIG.
Sources
- cisa.govhttps://www.cisa.gov/news-events/alerts/2026/11/18/cisa-adds-four-known-exploited-vulnerabilities-catalog
- cisa.govhttps://www.cisa.gov/known-exploited-vulnerabilities-catalog
- securityonline.infohttps://securityonline.info/cisa-adds-4-exploited-flaws-kev-synology-litellm/
- shattered.iohttps://shattered.io/litellm-axis-redcap-synology-dsm-four-flaws-join-cisa-kev-nov-18/
- securityaffairs.comhttps://securityaffairs.com/203260/hacking/google-gtig-warns-ai-agents-automate-exploits-steal-credentials-and-scale-cybercrime.html