BerriAI LiteLLM: CVE-2026-40217 RCE exploited (CISA KEV)
CISA confirms exploitation of CVE-2026-40217 in LiteLLM: code execution via the /guardrails/test_custom_code URI. Technical details and what to do now.

On 18 November 2026 the US CISA added CVE-2026-40217 to its Known Exploited Vulnerabilities (KEV) catalogue. The flaw affects BerriAI LiteLLM, the project that provides a single interface to the language models of many providers. According to the description in the National Vulnerability Database (NVD), it allows arbitrary code execution on the server through the /guardrails/test_custom_code URI. Inclusion in the catalogue means CISA has evidence of real-world exploitation.
This article is based on a single source: CISA’s entry in the KEV catalogue, with the NVD data. The entry does not say who is exploiting the flaw, against which organisations, since when or with which tools.
What the KEV entry says
- Name: BerriAI LiteLLM Unprotected Alternate Channel Vulnerability.
- Added to the catalogue: 18 November 2026.
- Deadline for US federal agencies: 2 December 2026.
- Use in ransomware campaigns: unknown («Unknown»).
- Required action: apply mitigations in line with the vendor’s instructions, complying with CISA’s BOD 26-04 directive on prioritising security updates based on risk and with its «Forensics Triage Requirements»; for cloud services follow the same directive, or stop using the product if mitigations are not available. CISA also asks organisations to assess the internet exposure of every asset.
The reference to forensic triage requirements means that patching is not enough: you also need to check whether the instance has already been compromised.
Technical details
According to the NVD, LiteLLM releases up to 8 April 2026 allow remote attackers to execute arbitrary code via bytecode rewriting sent to the /guardrails/test_custom_code URI. The entry does not describe the mechanism any further.
- Weaknesses: CWE-420, unprotected alternate channel, and CWE-94, improper control of generation of code (code injection).
- CVSS 3.1: 8.8 (high), with the same score from two assessors, including MITRE. Vector
AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H: network attack, low complexity, low privileges required, no user interaction, high impact on confidentiality, integrity and availability. - NVD entry status: published on 10 April 2026, last modified on 18 November 2026 («Modified»).
There is a difference worth noting: the description speaks of «remote attackers», while the CVSS vector indicates that low privileges are needed, meaning already authenticated access with a limited role. The entry does not say which role or type of key is sufficient.
Among its references, the NVD lists advisory X41-2026-001 by X41 D-Sec, classified as containing exploit details and mitigations, two Red Hat errata (RHSA-2026:24866 and RHSA-2026:30056), the CVE page and the report in Red Hat’s bugzilla. The flaw’s entry has therefore been public since 10 April; what is new is the confirmation of exploitation.
In MITRE ATT&CK terms, exploiting an instance reachable over the network maps to T1190 (Exploit Public-Facing Application). The entry does not include public indicators of compromise.
Affected versions
The NVD CPE configuration lists LiteLLM up to and including the 2026-04-08 release. The entry does not give a fixed version number: it must be taken from the vendor’s guidance, from the X41 D-Sec advisory or, if you use Red Hat packages, from the errata listed. Installations that have not been updated since April should be considered vulnerable.
What it means for European companies
LiteLLM often sits at the centre of how a company uses language models: internal applications call it instead of the individual providers. Code execution on that server puts at risk everything the process can read, including its configuration. The KEV catalogue is binding only on US federal agencies, but it is a widely used reference for setting patching priorities. For organisations subject to NIS2 (in Italy, Legislative Decree 138/2024), which must manage vulnerabilities based on risk, a flaw with confirmed exploitation should be handled with high priority; the same applies to healthcare organisations that use these tools on texts containing patient data.
What to do now
- Inventory: identify every LiteLLM instance, including those started by development teams, in containers or in test environments, and check its version.
- Update: install the fixed version indicated by the vendor or, on Red Hat systems, the updates in errata RHSA-2026:24866 and RHSA-2026:30056.
- Exposure: as CISA requests, assess whether the instance is reachable from the internet or from untrusted networks. If it does not need to be, close it off; until you can update, consider blocking requests to
/guardrails/test_custom_codeat the reverse proxy. - Access: since low privileges are enough, review the users and keys that can access the instance and remove those no longer needed.
- Compromise assessment, as CISA requires with forensic triage: search the proxy and web server logs for requests to
/guardrails/test_custom_codeand check the server for unusual processes, files and outbound connections. - If you find suspicious signs, treat the event as an incident: updating does not remove access already obtained, and keys and secrets in the configuration should be considered exposed and replaced.
How ARXDOME helps
ARX Expose shows what can be seen of your company from outside, including exposed services and known vulnerabilities on public services: a LiteLLM instance reachable from the internet shows up there. ARX Shield brings into the portal the open vulnerabilities per machine and per package, highlighting those already exploited by attackers. With ARX Scout Assessment, once authorised, the non-destructive vulnerability assessment flags outdated software and exposed administration panels.
Last updated:


