Skip to content
Blog

BerriAI · LiteLLM · Vanderbilt

LiteLLM and REDCap Added to CISA KEV: Two Remote Code Execution Flaws (CVE-2026-40217, CVE-2026-90817)

Added to CISA KEV on November 18: LiteLLM (CVE-2026-40217, 8.8) via /guardrails/test_custom_code and REDCap (CVE-2026-90817, 9.8) via public surveys.

4 min readCVE-2026-40217CVE-2026-90817

On November 18, 2026, CISA added four vulnerabilities to the KEV catalog. Two of them hit self-hosted web applications that often live outside what the security team actually monitors:

  • LiteLLM (BerriAI), the gateway that puts a single API in front of many LLM providers;
  • REDCap (Vanderbilt University), the research data capture platform widely used in hospitals and universities.

Both lead to code execution on the server. The deadlines differ: December 2 for LiteLLM, November 21 for REDCap.


The Two Vulnerabilities

LiteLLMREDCap
CVECVE-2026-40217CVE-2026-90817
CVSS 3.18.8 (HIGH) — MITRE and a second CNA, identical values9.8 (CRITICAL) — Securifera (CNA)
VectorAV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:HAV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
CWECWE-420, CWE-94CWE-73, CWE-94
Authenticationrequired (low privileges, PR:L)none, but a valid public survey hash is needed
NVD publication2026-04-102026-09-20
NVD statusModifiedUndergoing Analysis
Added to CISA KEV2026-11-182026-11-18
CISA deadline2026-12-022026-11-21

Neither score was computed by NVD itself: both come from CNAs.


LiteLLM: CVE-2026-40217

NVD description:

LiteLLM through 2026-04-08 allows remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI.

As its name suggests, /guardrails/test_custom_code is there to test custom guardrail code. According to NVD, an attacker can use it to run arbitrary code on the server through bytecode rewriting. The CWE-420 classification (Unprotected Alternate Channel) describes a side channel that slips past the intended controls.

The vector says PR:L: you need authenticated access, even low-privileged. On an LLM gateway shared across teams, where many people and apps hold a key, that bar is low.

Versions: NVD lists releases through 2026-04-08. It gives no fixed version number: check GitHub advisory GHSA-wxxx-gvqv-xp7p, cited by CISA, and the X41 D-Sec advisory (x41-2026-001). Red Hat has also published errata (RHSA-2026:24866, RHSA-2026:30056) for products that bundle it.

Seven months passed between the NVD publication (April) and the KEV listing. It's the third LiteLLM flaw CISA has added to the catalog this year, after CVE-2026-42208 in May and CVE-2026-59822 in September. An LLM gateway holds the API keys for every model provider an organization uses — a valuable target.


REDCap: CVE-2026-90817

NVD description (excerpt):

An unauthenticated Remote Code Execution vulnerability was found in the survey passthrough routing and Data Import processing logic […] The attacker does not have to be authenticated in order to exploit this, but exploitation requires knowledge of a valid public survey hash. This vulnerability exists in REDCap 13.3.0 and higher.

Two defects chain together:

  1. survey passthrough routing lets a request from a public survey context reach a controller route it was never meant to reach;
  2. the data import logic accepts a crafted file-path or stream parameter (CWE-73, external control of a file path), which leads to code execution (CWE-94).

The only precondition is knowing a valid public survey hash. But a public survey is meant to be shared: its link goes out by email, on websites, in QR codes. In practice — this is our inference, not something the CVE states — that condition won't protect an instance that runs public surveys.

Versions: REDCap 13.3.0 and later, per the description. NVD hasn't published a CPE configuration or a fixed version yet. Check the REDCap community site cited by CISA (projectredcap.org) and the Securifera advisory.


Detection

LiteLLM

  • Requests to /guardrails/test_custom_code in proxy or application logs, especially from keys that have no reason to test guardrails
  • Unusual child processes spawned by the LiteLLM Python process
  • Abnormal use of model-provider API keys (volume, models called)

REDCap

  • Requests from public survey contexts to data import routes
  • Unknown PHP files in the web root
  • Data imports nobody started

Mitigation

LiteLLM

  1. Upgrade to a release that includes the fix, per the GHSA advisory
  2. Meanwhile, block /guardrails/test_custom_code at the reverse proxy if you don't use the feature
  3. Don't expose the LiteLLM admin UI or API to the Internet
  4. If exploitation is possible, rotate every API key stored in the gateway

REDCap

  1. Apply the fixed release published by REDCap before November 21
  2. Review active public surveys and close the ones no longer needed
  3. Check server file integrity, and follow the forensic triage requirements of directive BOD 26-04 that CISA points to
  4. A REDCap instance often holds sensitive research data, sometimes health data: if it was compromised, bring in your data protection officer

Why Continuously Monitor Your Self-Hosted Apps

Neither LiteLLM nor REDCap shows up in the usual network inventory. One gets installed by a data team, the other by a clinical research unit, and security finds out they exist when CISA adds them to KEV.

With cveo.tech, track these applications with their versions too, and get an alert as soon as a CVE affecting them lands in the CISA KEV catalog.

Know which CVEs hit the versions you run.

Add the product and its version: cveo.tech lists the CVEs that affect it right away, then alerts you by email when a new one affects it — including before NVD has finished its analysis. Free, an account is all it takes.

Every Monday

The week's critical CVEs, in your inbox

One email a week: the CVSS ≥ 9 vulnerabilities published in the last seven days, plus our latest analyses. Nothing else.

Double opt-in by email. Unsubscribe in one click, any time.