LiteLLM has a sandbox escape in its custom-code guardrail test endpoint that can allow arbitrary code execution in the proxy process, according to GitHub advisory GHSA-wxxx-gvqv-xp7p. The advisory says reaching the endpoint requires a proxy-admin credential in default configurations, and its advice is: “Upgrade to 1.83.11 or later.” On November 18, 2026, the Cybersecurity and Infrastructure Security Agency (CISA) added the flaw, tracked as CVE-2026-40217, to its Known Exploited Vulnerabilities (KEV) catalog, “based on evidence of active exploitation.”

#What changed?

The GitHub advisory, titled “Sandbox escape in custom-code guardrail,” describes a flaw in LiteLLM’s custom-code guardrail test endpoint, POST /guardrails/test_custom_code. It says the “endpoint runs user-supplied Python inside a hand-rolled sandbox,” “allowing arbitrary code execution in the proxy process,” “which runs as root in the default Docker image.” It also says: “Reaching the endpoint requires a proxy-admin credential in default configurations.” The advisory page shows a publish date of May 7, 2026. The National Institute of Standards and Technology’s National Vulnerability Database (NVD) record for CVE-2026-40217 shows the CVE was published on April 10, 2026.

CISA’s alert dated November 18, 2026 said: “CISA has added four new vulnerabilities to its Known Exploited Vulnerabilities (KEV) Catalog”, and that this was “based on evidence of active exploitation.” It lists the LiteLLM item as “CVE-2026-40217 BerriAI LiteLLM Unprotected Alternate Channel Vulnerability.” CISA’s KEV catalog entry records the same date added, November 18, 2026, and points at the GitHub advisory. The Hacker News reported the same listing on November 19, 2026.

GitHub rates the advisory “Moderate,” with a “5.9 CVSS overall score” under CVSS v4. NVD lists a CVSS 3.1 base score of 8.8, labeled high. The Hacker News also cites a CVSS score of 8.8.

#Who is affected?

The advisory lists “Affected versions >= 1.81.8, < 1.83.10” and “Patched versions >= 1.83.10.” Its text says “Fixed in 1.83.11” and “Upgrade to 1.83.11 or later.” A references line lists “Patched release: v1.83.10-stable.” The advisory also says: “The hand-rolled sandbox has been replaced with RestrictedPython.”

Our analysis: The advisory’s patched-versions field starts at 1.83.10, while its own upgrade text names 1.83.11. The advice to follow is the vendor’s sentence: “Upgrade to 1.83.11 or later.”

The KEV entry sets a due date of December 2, 2026. That deadline applies to US federal civilian agencies under Binding Operational Directive 26-04 (BOD 26-04), according to the KEV entry and NVD. CISA’s November 18 alert says: “While BOD 26-04 applies only to FCEB agencies, CISA encourages all organizations to adopt risk-based vulnerability management and prioritize remediation of KEV Catalog vulnerabilities”. The KEV entry also says to “Apply mitigations in accordance with vendor instructions”, “or discontinue use of the product if mitigations are unavailable.”

#What should operators do?

Anyone running an affected LiteLLM version should upgrade. The GitHub advisory’s advice is: “Upgrade to 1.83.11 or later.” If an immediate upgrade is not possible, the workaround is: “If upgrading is not immediately possible, block POST /guardrails/test_custom_code at your reverse proxy or API gateway.”

Operators choosing a later LiteLLM release should also review our earlier LiteLLM supply-chain coverage before they pin a build. A follow-up on that incident is in our report on the LiteLLM cosign key update.

#What does Google say?

The Hacker News ties CISA’s listing to a Google Threat Intelligence Group report. That Google Cloud blog post, dated November 3, 2026, is Google’s account of LLMJacking intrusions. It should be read as Google’s account, and it differs from the GitHub advisory on what CVE-2026-40217 is.

Google’s researchers write: “Mandiant responded to multiple incidents in 2026 in which threat actors hijacked enterprise access to Large Language Models (LLMs), a tactic known as ‘LLMJacking.’” They say “these intrusions impacted organizations in North America, South America, and Asia Pacific across the financial services, retail, technology, and transportation sectors.” They also write: “Across individual incidents, we observed fraudulent charges associated with resource consumption exceeding $100,000 USD in a single day, with cumulative charges at some organizations reaching $1 million USD before access was revoked.” Those dollar figures are about LLMJacking incidents in general, not about this LiteLLM flaw alone.

On the CVE itself, Google’s post says: “The actor exploited a user ID rebind vulnerability in the /key/regenerate endpoint (CVE-2026-40217) to elevate standard Virtual Keys into full PROXY_ADMIN master administrative credentials, establishing deep, persistent control over the compromised environment.” Google also says: “we have observed instances of threat actors targeting vulnerabilities specifically within LLM routing software to escalate privileges.” Google’s defensive advice is to extend vulnerability management to centralized model proxy and routing gateways, including LiteLLM, and to patch privilege-escalation and authentication-bypass flaws such as CVE-2026-40217.

The GitHub advisory, CISA’s KEV entry and NVD all describe CVE-2026-40217 as a sandbox escape in the custom-code guardrail test endpoint, which needs a proxy-admin credential in default configurations. Google’s November 3, 2026 post attaches the same CVE number to a different bug: a user ID rebind in the /key/regenerate endpoint that, in Google’s account, turned ordinary Virtual Keys into PROXY_ADMIN credentials. Those starting points are not the same, and the sources should not be read as a single attack chain.

Google’s broader reporting on LLM-related threats is also covered in our article on the GTIG AI Threat Tracker.

#How does it compare?

CISA’s KEV catalog is a list of vulnerabilities the agency says have been exploited. An earlier example on this site is CISA’s KEV listing of a Cisco BroadWorks flaw. The LiteLLM item is a separate listing, added on November 18, 2026.

Our analysis: Operators do not need to resolve the clash between Google’s description and the GitHub advisory before they act. The advisory’s fix and workaround are specific: upgrade to 1.83.11 or later, or block POST /guardrails/test_custom_code until that upgrade is in place. Anyone moving to a later release should confirm they are installing a build they trust.

#What is still unclear?

The sources leave open how the two descriptions of CVE-2026-40217 relate, and which patched version is the first fixed release.

  • How do Google’s description of CVE-2026-40217, a user ID rebind in the key-regeneration endpoint, and the GitHub advisory’s sandbox escape in the guardrail test endpoint relate?
  • Which release should operators treat as the first patched version, 1.83.10 or 1.83.11?
  • In which incidents, and at what scale, was CVE-2026-40217 used? The KEV entry lists known ransomware campaign use as unknown.