Skip to content
TEN Brief Ten verified stories a day 2026.11.20 KO

이 기사는 한국어로도 읽을 수 있습니다 →

Tech · 16 min read · Explainer

LiteLLM CVE-2026-40217 patch — 1.83.11, but go to 1.83.14

The fixed version for CVE-2026-40217 in LiteLLM is stated two ways inside the primary sources themselves. The LiteLLM security advisory (GHSA-wxxx-gvqv-xp7p) gives an affected range of 1.81.8 up to but not including 1.83.10, while the text of the same advisory says the fix is in 1.83.11 and tells users to upgrade to 1.83.11 or later. The NVD record and the CISA catalog entry carry no version number. Obsidian Security, which reported this flaw together with two privilege-escalation flaws (CVE-2026-47101 and CVE-2026-47102), says the whole chain is closed only in v1.83.14-stable. The practical floor is therefore 1.83.14 or later. The US Cybersecurity and Infrastructure Security Agency (CISA) added the flaw to its Known Exploited Vulnerabilities (KEV) catalog on November 18, 2026 and gave federal civilian agencies until December 2. Exploitation needs a proxy-admin credential. Where an upgrade has to wait, block POST /guardrails/test_custom_code at the reverse proxy.

A platform engineer seen from behind checking systems at a standing desk with a laptop, a monitor and a small server rack in a sunlit office

The three lines

  • Versions — the advisory range says affected below 1.83.10, the advisory text says upgrade to 1.83.11 or later, and the privilege-escalation chain is closed in v1.83.14-stable. 1.83.14 or later clears all three. 1.107.0 is an unrelated malicious release and should be avoided
  • Conditions — a proxy-admin credential (typically the master key) is needed to reach the endpoint. Before 1.83.14, separate flaws let an ordinary account become admin. Library-only use has no such endpoint
  • Status — disclosed April 8, 2026; fixed in mid-April; added to CISA KEV November 18; federal deadline December 2. LLMjacking claims come from two secondary reports, and the original report was not found

Key questions

LiteLLM CVE-2026-40217 patch version
**1.83.14 or later clears all three thresholds that different sources give (1.83.10, 1.83.11, 1.83.14).** | Source | Fixed version stated | Kind | |---|---|---| | LiteLLM advisory header (affected range) | 1.83.10 (1.81.8 up to but not including 1.83.10 affected) | Primary | | LiteLLM advisory text | 1.83.11 (upgrade to 1.83.11 or later) | Primary | | Reference link in the same advisory | v1.83.10-stable | Primary | | OSV / PyPA record (PYSEC-2026-2601) | 1.83.10 | Mirror of the advisory | | NVD, CISA KEV | No version number | Primary | | X41 D-Sec advisory | Patched version still TBD | Primary | | Obsidian Security, CSA | v1.83.14-stable (whole chain) | Finder / secondary |
Does CVE-2026-40217 require authentication
**Yes. The endpoint is reachable only with a proxy-admin credential, usually the master key.** | Item | Detail | |---|---| | Location | POST /guardrails/test_custom_code on the LiteLLM proxy | | Type | Escape from the guardrail test sandbox, leading to code execution | | Privileges required | Proxy admin (LiteLLM advisory); master key (X41) | | CVSS 3.1 | 8.8 (CVE record, Red Hat) | | CVSS 4.0 | 8.7 (X41), 5.9 (LiteLLM repository advisory) | | Result | Code execution as the proxy process, which is root in the default Docker image | | Library-only use | No such endpoint unless the proxy is running |
CVE-2026-40217 CISA KEV due date
**December 2, 2026. CISA set it for US federal civilian agencies when it added the flaw to KEV on November 18, 2026.** | Date | Event | |---|---| | February 5, 2026 | 1.81.8, the first release with the endpoint, is published | | April 8 | X41 D-Sec publishes its advisory with no fix available | | April 10 | CVE record published | | April 19, 22, 26 | 1.83.10, 1.83.11, 1.83.14 uploaded to PyPI | | November 18 | CISA adds the flaw to KEV | | December 2 | Federal civilian agency deadline |

The fixed version for CVE-2026-40217 in LiteLLM is given three different ways: 1.83.10, 1.83.11 and v1.83.14-stable. The disagreement was not created by second-hand reporting. A single security advisory from BerriAI, the company behind LiteLLM, lists the affected range as "below 1.83.10" and then says in its text that the fix is "in 1.83.11." The third number comes from the researchers who reported this flaw together with two others and say the whole path is closed only in v1.83.14-stable. The short answer: 1.83.14 or later clears all three.

The US Cybersecurity and Infrastructure Security Agency (CISA) added the flaw to its Known Exploited Vulnerabilities (KEV) catalog on November 18, 2026, with a December 2 deadline for federal civilian agencies. The fix has existed since mid-April, so a flaw repaired seven months earlier has now been confirmed as used in real attacks. This article sets out what the primary sources actually say, what privileges an attacker needs, how to check a deployment and in what order to act, and how much evidence stands behind the exploitation reports. It is written for defenders and contains no exploit detail.

1. Which version is fixed — 1.83.10, 1.83.11 or 1.83.14, according to the primary sources

LiteLLM is an open-source tool that puts many AI model providers behind one API. It is used as a Python library and also as a "proxy" server that an organization runs internally to issue keys and track spending. The flaw is in the proxy.

SourceWhat it saysKindOpened directly?
LiteLLM advisory (GHSA-wxxx-gvqv-xp7p), headerAffected: 1.81.8 up to but not including 1.83.10Vendor, primaryYes
Same advisory, textFixed in 1.83.11; upgrade to 1.83.11 or laterVendor, primaryYes
Same advisory, reference linkLinks v1.83.10-stable as the patched releaseVendor, primaryYes
OSV / PyPA record (PYSEC-2026-2601)Introduced 1.81.8, fixed 1.83.10Mirror of the advisoryYes
NVD CVE recordNo version number; "LiteLLM through 2026-04-08"PrimaryVia the API
CISA KEV entryNo version number; links the vendor advisoryPrimaryYes
X41 D-Sec advisoryPatched version still "TBD"Finder, primaryYes
Release notes, v1.83.10-stable and v1.83.14-stableNo mention of the sandbox or the CVEVendor, primaryYes
Obsidian SecuritySandbox replacement shipped in v1.83.10-stable; whole chain closed in v1.83.14-stableFinder, primaryYes
HOLSays the sandbox files in 1.83.9 through 1.83.11 are identicalSecondary, single sourceYes

Here is how to read those rows.

1.83.10 is the number implied by the vendor advisory's affected range and by the OSV record that mirrors it. The fix itself is a change merged on April 15, 2026 (PR #25818) that replaces a home-made, string-matching sandbox with RestrictedPython, an established library. 1.83.10 was uploaded to PyPI on April 19.

1.83.11 is what the text of the same advisory recommends. The advisory does not say why the two numbers differ. HOL reports comparing the distribution files and finding the relevant files identical in 1.83.9, 1.83.10 and 1.83.11; if that is right, the gap is one of labeling rather than code. That comparison comes from one source.

v1.83.14-stable is not the fix for this flaw alone. In February 2026 Obsidian Security reported three problems to LiteLLM at once: one that lets an ordinary account reach admin-only routes (CVE-2026-47101), one that lets a user assign themselves the admin role (CVE-2026-47102), and this flaw, which turns admin access into code execution. According to Obsidian, v1.83.10-stable carried the role-escalation fix and the sandbox replacement, and the remaining bypasses were closed in v1.83.14-stable. A Cloud Security Alliance (CSA) research note from June recommends the same release.

That makes the practical decision straightforward.

Installed versionCVE-2026-40217Escalation chain (47101, 47102)Verdict
1.81.7 or earlierEndpoint does not existCheck other flaws separatelyNot in scope for this flaw, but an old release
1.81.8–1.83.9AffectedOpenAct immediately
1.83.10Fixed by the advisory range; short of the advisory text's floorPartly fixed (Obsidian)Upgrade
1.83.11–1.83.13FixedSome bypasses remain (Obsidian)Upgrade
1.83.14 or laterFixedClosed (Obsidian)Meets all three thresholds
1.107.0——Unrelated malicious release; do not use

As checked on November 20, 2026, the latest release on PyPI is 1.110.1. Unless something pins a deployment to April-era builds, the current stable release is the right target. The KEV catalog already holds three other LiteLLM flaws: a SQL injection added on May 8 (CVE-2026-42208), a command injection added on June 8 (CVE-2026-42271) and an authentication flaw added on September 2 (CVE-2026-59822). 1.83.14 is the floor for this flaw only; it does not cover everything found since.

2. What breaks — the admin key is required, and it yields the whole server

The LiteLLM proxy has a "guardrails" feature: rules that inspect requests going to a model and responses coming back. An administrator can write a rule in Python, and the endpoint for trying that code out is POST /guardrails/test_custom_code. Because it runs code supplied by someone else, it had a sandbox meant to block dangerous operations. CVE-2026-40217 is the failure of that sandbox. What a sandbox is, and why a single open channel defeats one, is covered in the October 6, 2026 article 「샌드박스란 — AI 에이전트를 가둔 상자는 열어 둔 통로 하나로 뚫린다」 (an explainer on sandboxes for AI agents).

ItemDetailSource / meaning
CVECVE-2026-40217Published April 10, 2026 (NVD)
LocationPOST /guardrails/test_custom_code on the proxyEndpoint introduced in 1.81.8, the start of the advisory range
TypeSandbox escape leading to code executionCWE-420 in KEV, CWE-913 in the advisory, CWE-94 per X41 and Red Hat
CauseA hand-rolled sandbox built on string checksThe advisory says only that it can be escaped with bytecode-level techniques
Privileges requiredA proxy-admin credentialThe advisory qualifies this as "in default configurations"; X41 says the master key
CVSS 3.18.8 (High)CVE record (MITRE), Red Hat
CVSS 4.08.7 (High)X41 assessment
CVSS 4.05.9 (Moderate)LiteLLM repository advisory, which rates attack complexity High and privileges required High
ResultArbitrary code execution as the proxy processRoot in the default Docker image (advisory, X41)
Found byMarkus Vervier of X41 D-SecFound February 13, 2026; published April 8
Added to KEVNovember 18, 2026Due December 2; ransomware use Unknown

The spread between 8.8 and 5.9 comes down to how "privileges required" is scored. X41 and the CVE record rate it Low; LiteLLM rates it High. Both agree an admin key is needed. They disagree on how hard that key is to get. How to read these scores is covered in the October 6, 2026 article 「CVSS란 — 깃랩 취약점이 10점이 아니라 9.9점인 이유」 (an explainer on CVSS).

A fair objection is that an admin key already allows everything. The difference is scope. The admin key opens the proxy's features; this flaw hands over the server the proxy runs on.

An attacker with only the admin keyAn attacker who also uses this flaw
Uses the proxy's management APIRuns commands on the server, as root in the default image
Issues and deletes keys, changes model settingsReads environment variables and config files, where provider API keys live
Sends requests through the proxyMoves to other systems on the same network
Is locked out once the key is rotatedLeaves things on the server that survive a key rotation

And the "admin key required" condition was weaker than it sounds on anything before 1.83.14. In the chain Obsidian described, the starting point is not an administrator but a default low-privilege account (internal_user). Two steps make that account an admin; the third step is this flaw. Obsidian rates the chain 9.9. In an organization that has handed LiteLLM keys to dozens of developers, one leaked key could have been enough.

Comparing the flaw with the three others added to KEV on the same day makes its character clearer.

CVEProductAuthenticationDue dateForensic triage required
CVE-2026-40217BerriAI LiteLLMRequired (admin)December 2, 2026No
CVE-2024-10441Synology DSMRemote attacker, per the KEV descriptionNovember 21, 2026Yes
CVE-2026-90436Apache AxisNot requiredNovember 21, 2026Yes
CVE-2026-90817REDCapNot required (needs a public survey link)November 21, 2026Yes

LiteLLM is the only one of the four with a 14-day deadline; the others have three days. According to the CISA alert, the governing directive (BOD 26-04) calls for fast remediation of high-risk flaws on internet-exposed assets that give total control when exploited. A flaw that requires authentication appears to have been placed one tier below that, although CISA has not explained the classification.

3. What to do — check the version, upgrade, and block the endpoint where an upgrade must wait

The order is: establish whether you are affected, upgrade, block on anything not yet upgraded, look for calls to the endpoint, then decide on key rotation.

StepTaskHowWhat to look for
1Confirm the proxy is in useIs LiteLLM running as a server, or only imported as a library in code?Library-only use has no such endpoint
2Version in a Python environmentpip show litellmVersion 1.83.14 or later
3Version inside a containerdocker exec <container> pip show litellmSame. Check the installed version, not the image tag
4Image tagdocker inspect --format '{{.Config.Image}}' <container>For a moving tag such as main-latest, find out when it was pulled
5Upgradepip install --upgrade "litellm[proxy]"; for Docker, pull a new image and recreate the containerRe-run steps 2 and 3
6Look for callsSearch reverse-proxy access logs and container logs for test_custom_codeCalls when no administrator was testing guardrails are a red flag
7Decide on key rotationIf step 6 turns up calls nobody can explain: the master key and every provider keySee below

Steps 3 and 4 are separate for a reason. A tag such as main-latest points at whatever version existed when it was pulled. X41 confirmed the flaw on the main-latest image. An image pulled before April under that same tag name is a vulnerable build. Pulling a new image also does not change a container that is already running; Ostorlab notes the container has to be recreated so the new package is what actually runs.

For anyone pinning versions at step 5, avoid 1.107.0. That is the malicious release published to PyPI and Docker on October 21, 2026, and this article confirmed it is no longer listed on PyPI. That incident was a supply-chain compromise of the release path; this one is a defect in legitimate code. Causes and remedies differ. The earlier incident is covered in the October 26, 2026 article 「LiteLLM 1.107.0 악성 버전 확인법 — 136일 방치된 마스터 키 하나가 배포 파이프라인을 열었다」 (on identifying the malicious 1.107.0 release).

Where an upgrade cannot happen yet, the vendor advisory names one interim measure: block POST /guardrails/test_custom_code at the reverse proxy or API gateway. The rest are exposure-reduction steps from the X41 advisory and from security vendors.

MeasureWhat it doesRecommended byLimit
Block the endpointReject POST /guardrails/test_custom_code at the edgeLiteLLM advisory, X41, Ostorlab, HOLThe old sandbox is still there; does not repair a compromised server
Separate the admin pathMake the management API and admin UI reachable only from internal networks or a VPNX41, HOL, CSAStill open to an attacker already inside
Control the master keyDo not give the master key to ordinary API consumers; remove it from shared documents and repositoriesX41, HOLA key that has already leaked must be rotated
Run as non-rootRun the container as an unprivileged userX41, HOL, CSALimits the damage; does not prevent it
MonitorLog calls to the endpoint and unexpected child processes of the proxyHOL, CSADetection, not prevention

Blocking the endpoint does not affect normal service. It is used only when testing guardrail code; chat and completion requests take other routes (HOL).

Key rotation is widely advised, but it is not automatically the first step. Upgrade or block first; rotating a key while the hole is open puts the new key at the same risk. If step 6 turns up unexplained calls, the picture changes. HOL advises treating the proxy as compromised in that case: rebuild from a clean image and rotate the master key and every provider key stored in the proxy. If code ran on the server, the keys in its environment should be assumed read.

How the master key is rotated depends on configuration, according to the LiteLLM documentation. If a separate encryption key (LITELLM_SALT_KEY) is set, the master key is only an authentication credential: change the value and restart every instance. If it was never set, the master key also encrypts credentials stored in the database, and a re-encryption procedure is needed. The documentation says to back up the database first either way.

4. How strong the exploitation evidence is — CISA's listing is confirmed; LLMjacking is second-hand

LLMjacking means stealing someone else's AI service credentials, running inference on them, and leaving the victim with the bill. The mechanics are covered in the October 18, 2026 article 「API 키 도용이란 — 훔친 AI 키로 해킹 도구를 돌리고 요금은 키 주인이 낸다」 (an explainer on API key theft). A LiteLLM proxy gathers an organization's provider keys in one place, which is why it is often named as a target. Whether this particular flaw was used for that purpose needs to be separated by source.

WhoWhat they saidStrength
CISAAdded to KEV on November 18, 2026 on evidence of exploitation. No attacker, victim or method named. Ransomware use UnknownPrimary; confirmed
NVDCVE record updated with the KEV date and the December 2 due datePrimary; confirmed
X41 D-SecApril 8 advisory includes a proof of concept. No mention of exploitation. Page not updated sincePrimary; confirmed
Obsidian SecurityPublished the chain on June 11. No statement on exploitation in the wildPrimary; confirmed
Security AffairsReports that Google Threat Intelligence Group (GTIG) said attackers escalated limited inference keys to admin credentials and then ran LLMjacking, against financial, retail, transportation and technology firms in North America, South America and Asia-PacificSecondary; GTIG original not found
MalloryRelays one blog (Sh3llc0d3) alleging LLMjacking, credential theft and cryptomining, and states the claims are not independently substantiatedSecondary; original not found
Ostorlab, HOLSay no victims or indicators of compromise are public beyond the KEV listingSecondary

In short, CISA has confirmed that someone used the flaw. That it was used for LLMjacking is reported by two outlets, and this article could not locate the underlying report. Mallory itself says the claims it relays are unverified. The method Security Affairs describes, raising a limited key to admin and then using the flaw, has the same shape as the chain Obsidian published in June. No source confirms that real attacks used those two flaws (CVE-2026-47101 and CVE-2026-47102).

With no indicators of compromise published, defenders are left with behavior: calls to the endpoint, unfamiliar child processes of the proxy, and the provider invoices. Usage at unusual hours, calls to models nobody uses, and a jump in token counts are the fastest sign that a key has leaked.

DateEventSource
February 5, 20261.81.8 uploaded to PyPI, the first release with the endpointPyPI, advisory range
February 13X41 D-Sec finds the flawX41 advisory
February 18LiteLLM acknowledges the reportX41 advisory
February 19Obsidian reports all three flawsObsidian
April 8Embargo expires; X41 publishes with no fix availableX41 advisory
April 10CVE-2026-40217 publishedNVD
April 15Sandbox replacement (PR #25818) mergedGitHub
April 191.83.10 uploaded to PyPIPyPI
April 221.83.11 uploaded to PyPIPyPI
April 261.83.14 uploaded to PyPIPyPI
May 7–11LiteLLM security advisory published; GitHub Advisory Database entryAdvisory, OSV
June 11Obsidian publishes the escalation chainObsidian
November 18CISA adds the flaw to KEVCISA alert, KEV catalog
December 2Federal civilian agency deadlineKEV catalog, NVD

Two intervals stand out. The first is the 11 days from April 8 to April 19. X41 reported in February and followed up several times; when no fix arrived within the coordination window, it published without one. For that period a public proof of concept existed and a fixed release did not. The second is the 213 days from April 19 to November 18. Servers left unpatched for seven months after the fix became the targets. How a published fix gives attackers a lead is the subject of the November 18, 2026 article 「패치 디핑이란 — 패치가 공격 지도가 되는 이유, 악용까지 63일이 −7일로」 (an explainer on patch diffing); here attackers did not even need that, because the method was public before the fix.

The December 2 deadline binds only US federal civilian agencies. Organizations elsewhere, including in South Korea where this publication is based, have no obligation under it. KEV is nonetheless a list built on evidence of actual use, which makes it a strong input for ordering a patch queue.

5. Frequently asked questions

The code only does import litellm. Is it affected? The endpoint lives on the proxy server. HOL says library-only use without the proxy is not affected. Another team may still be running a proxy from the same package, so the first check is whether a LiteLLM proxy is running anywhere in the organization.

Is 1.83.10 safe? It is outside the advisory's affected range and below the floor the advisory text recommends (1.83.11 or later). According to Obsidian, bypasses that let an ordinary account reach admin routes were closed only in 1.83.14. Move to 1.83.14 or later.

If an admin key is required, is this really urgent? Master keys tend to end up in environment variables, deployment scripts and internal documents. The LiteLLM supply-chain incident of October 2026 began with one master key committed to a public repository. And before 1.83.14 there was a separate route from an ordinary key to admin. CISA's confirmation of exploitation is evidence that the condition is being met in practice.

Is blocking the endpoint enough? As an interim measure, yes; it is the one the vendor advisory names. The old sandbox code remains, though, and the three other LiteLLM flaws on KEV are untouched by it. It buys time until an upgrade.

What about a managed or hosted LiteLLM? Whoever operates it needs to confirm the version. This article did not verify the status of individual managed services.

Is the score 8.8 or 5.9? Different scale versions and different assessors. 8.8 is CVSS 3.1 in the CVE record, 8.7 is X41's CVSS 4.0 score, and 5.9 is the CVSS 4.0 score LiteLLM gave in its own advisory. The gap reflects how heavily the admin-key condition is weighted. The KEV listing is a better guide to priority than any of the numbers.

Does the December 2 deadline apply outside the US government? Not legally. Given confirmed exploitation and a target that concentrates an organization's AI keys, it belongs near the front of the queue regardless.

6. What remains open and what could not be confirmed

ItemConfirmedNot confirmed
Fixed versionAdvisory range 1.83.10; advisory text 1.83.11; Obsidian's chain closure in v1.83.14-stableWhy the two numbers in the advisory differ; no BerriAI clarification was found
Code difference across 1.83.9–1.83.11HOL's statement about its file comparisonNot compared directly; single source
Release datesPyPI upload times (April 19, 22, 26)GitHub release pages and Obsidian give dates a few days off
ScoreCVSS 3.1 8.8; CVSS 4.0 8.7 (X41) and 5.9 (LiteLLM advisory)Whether 5.9 or the 7.5 reported for the GitHub Advisory Database is current
ExploitationCISA KEV listing on November 18Attackers, number of victims, indicators of compromise, ransomware use
LLMjackingThat Security Affairs and Mallory reported itThe GTIG report and the blog post; the two differ on targeted sectors
Use of the chain in attacksThat Obsidian published the chain in JuneWhether real attacks used CVE-2026-47101 and CVE-2026-47102
DeadlineDecember 2 in the KEV catalog and NVDThe CISA alert body states no date
Why 14 daysThe other three flaws that day got three days and a forensic triage requirementCISA has not explained the classification
Vendor positionThe May security advisoryAny BerriAI statement after the KEV listing

The order, then: find every server running LiteLLM as a proxy. Check the installed version with pip show litellm; if it is below 1.83.14, move to the current stable release and recreate the container. Where that cannot happen yet, block POST /guardrails/test_custom_code at the edge. Search the logs for calls to that endpoint, and if any cannot be explained, rebuild the server and rotate the master key and the provider keys.

The separate incident in which a malicious release was published through a hijacked release path is covered in the October 26, 2026 article 「LiteLLM 1.107.0 악성 버전 확인법 — 136일 방치된 마스터 키 하나가 배포 파이프라인을 열었다」. How sandboxes work is in the October 6, 2026 article 「샌드박스란 — AI 에이전트를 가둔 상자는 열어 둔 통로 하나로 뚫린다」, how stolen AI keys are used is in the October 18, 2026 article 「API 키 도용이란 — 훔친 AI 키로 해킹 도구를 돌리고 요금은 키 주인이 낸다」, and what the score means is in the October 6, 2026 article 「CVSS란 — 깃랩 취약점이 10점이 아니라 9.9점인 이유」.

Sources

  1. CISA — CISA Adds Four Known Exploited Vulnerabilities to Catalog (2026-11-18)
  2. CISA — Known Exploited Vulnerabilities Catalog JSON (catalog version 2026.11.18)
  3. BerriAI LiteLLM — Security advisory GHSA-wxxx-gvqv-xp7p: Sandbox escape in custom-code guardrail
  4. NVD — CVE-2026-40217 (API record, published 2026-04-10, modified 2026-11-19)
  5. OSV — PYSEC-2026-2601 / GHSA-wxxx-gvqv-xp7p records
  6. PyPI — litellm release listing (JSON, upload time per version)
  7. GitHub — BerriAI/litellm release v1.83.10-stable
  8. GitHub — BerriAI/litellm release v1.83.14-stable
  9. GitHub — BerriAI/litellm PR #25818: replace custom_code sandbox with RestrictedPython
  10. X41 D-Sec — Advisory X41-2026-001: Guardrail Sandbox Escape in LiteLLM (2026-04-08)
  11. Obsidian Security — Breaking LiteLLM: From Low-Privilege User to Admin and RCE (2026-06-11, updated 2026-09-07)
  12. Cloud Security Alliance — LiteLLM AI Gateway: KEV-Listed Attack Chain Enables Full Takeover (2026-06-17)
  13. LiteLLM docs — Master Key Rotations
  14. Red Hat — CVE-2026-40217
  15. Security Affairs — U.S. CISA adds Apache Axis, Vanderbilt REDCap, BerriAI LiteLLM and Synology DSM flaws to its KEV catalog (2026-11-19)
  16. Ostorlab — LiteLLM Guardrail Sandbox Escape RCE (CVE-2026-40217) (2026-11-19)
  17. HOL — CVE-2026-40217 LiteLLM guardrail sandbox escape on CISA KEV (2026-11-18)
  18. shattered.io — LiteLLM, Axis, REDCap, Synology DSM: four flaws join CISA KEV Nov 18
  19. Mallory — story on the LiteLLM guardrail sandbox escape reports (2026-11-19)

Verification

Published
Last modified
Cross-check
Checked against 19 independent sources.
Unverified
  • The LiteLLM advisory disagrees with itself on the fixed version. Its header gives an affected range ending below 1.83.10, its text says fixed in 1.83.11, and its reference link points to v1.83.10-stable. The advisory does not explain the difference, and no separate clarification from BerriAI was found. The release notes for v1.83.10-stable and v1.83.14-stable mention neither the sandbox fix nor the CVE. There is no v1.83.11-stable release page (404). PyPI lists 1.83.11 as uploaded on April 22, 2026.
  • HOL says it compared the PyPI distribution files and found the sandbox-related files identical in 1.83.9, 1.83.10 and 1.83.11, with 1.83.8 the last release carrying the old filter. This article did not compare the files itself. Single source.
  • That v1.83.14-stable closes the whole chain comes from Obsidian Security, the finder, and the Cloud Security Alliance note that relays it. The LiteLLM advisory does not mention 1.83.14. The blog (Sh3llc0d3) that Mallory cites as calling 1.83.14 the complete fix could not be located. The vendor advisories for CVE-2026-47101 and CVE-2026-47102 were not opened.
  • Release dates differ by source. PyPI upload times are April 19, 2026 for 1.83.10 and April 26 for 1.83.14. The GitHub release pages show April 27 for v1.83.10-stable and May 2 for v1.83.14-stable. Obsidian gives April 22 and April 25. The date tables in this article follow PyPI upload times (UTC).
  • Four scores are in circulation. The CVE record and Red Hat give CVSS 3.1 8.8; X41 gives CVSS 4.0 8.7; the LiteLLM repository advisory page read as CVSS 4.0 5.9 (Moderate). The GitHub Advisory Database record mirrored by OSV is marked HIGH with a different vector, and Ostorlab reports that score as 7.5. Which of 5.9 and 7.5 is current could not be settled.
  • The advisory has two publication dates. The LiteLLM repository advisory page read as May 7, 2026; the GitHub Advisory Database review and publication timestamp is May 11.
  • LLMjacking exploitation is reported by two secondary sources only. Security Affairs attributes it to Google Threat Intelligence Group (GTIG); Mallory relays a single blog's claims and says they are not independently substantiated. Neither the GTIG report nor that blog post was found. The two also differ on targeted sectors (Security Affairs adds transportation). No indicators of compromise are public.
  • The body of the CISA alert states no due date. December 2 is the dueDate value in the KEV catalog JSON and the cisaActionDue value in the NVD record. The KEV entry cites BOD 26-04; Security Affairs cites BOD 22-01. This article follows the KEV entry.
  • The NVD web page returned no body, so the NVD API record was used. It has no affected version range; the description says only LiteLLM through 2026-04-08.
  • The pip show command and the log searches in section 3 come from the HOL write-up. The Docker image tag check is a general Docker command. The statements that team and user keys cannot reach the endpoint and that library-only use is unaffected also rest on the HOL write-up alone, though they are consistent with the advisory's requirement of a proxy-admin credential.
  • The Hacker News article from June 2026 on the chain returned 403 and was not read.
Authoring
Reviewed by a person before publication. The full process is described in the Editorial.

Ten stories, once each morning

We send the three-line summaries only; the full pieces stay on the site. One-click unsubscribe, any time.

Related