LiteLLM's CVE-2026-40217 Is on CISA's KEV List, but Advisory and Google Name Different Endpoints
Security / news
LiteLLM's CVE-2026-40217 Is on CISA's KEV List, but Advisory and Google Name Different Endpoints
GitLab's advisory says the bug needs a proxy-admin credential, while Google's report says attackers used it to turn standard virtual keys into admin ones.
CISA added LiteLLM's CVE-2026-40217 to its Known Exploited Vulnerabilities catalog on Nov. 18 with a federal deadline of Dec. 2, but the three public descriptions of the bug disagree on which endpoint is vulnerable and how much access an attacker needs. Upgrade the LiteLLM proxy to at least 1.83.10, and to 1.83.14-stable if you want the fixes the Cloud Security Alliance says close the whole privilege chain.
The score is 8.8, and it overstates how easy the bug is on a default install, according to the advisory.
Three descriptions of one CVE
The GitLab advisory entry, dated May 11, covers versions "starting from 1.81.8 before 1.83.10" and rates the flaw 8.8 (AV:N/AC:L/PR:L/UI:N). It says POST /guardrails/test_custom_code "runs user-supplied Python inside a hand-rolled sandbox" that can be escaped with bytecode techniques. It also says: "Reaching the endpoint requires a proxy-admin credential" in default configurations.
The Hacker News quotes CISA's wording as remote code execution through bytecode rewriting at that same URI, and lists the federal due date as Dec. 2, 14 days after listing. The other three Nov. 18 entries carry a Nov. 21 deadline, three days out.
Google Threat Intelligence Group describes a different mechanism. Its Nov. 2 LLMjacking report says an 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."
| Source | Endpoint named | Attacker starts with |
|---|---|---|
| GitLab advisory (May 11) | /guardrails/test_custom_code | Proxy-admin credential |
| CISA entry via The Hacker News | /guardrails/test_custom_code | Remote, no stated prerequisite |
| Google GTIG report (Nov. 2) | /key/regenerate | Standard virtual key |
What Google reported, and what it left out
The GTIG report, written by nine named authors including Christopher Romano and Codi Starks, describes two clusters. One established initial access and passed on pre-validated virtual keys. A second used them against the victim's LiteLLM gateway. The report does not name the victim, the sector or the date of this intrusion, and it gives no request details.
Its wider figures are large: charges over $100,000 in a single day, cumulative charges at some organisations reaching $1 million, and one actor making more than 500,000 LLM API calls in under eight hours. Those figures cover the whole LLMjacking campaign, not only the LiteLLM intrusion.
- Charges in a single day, at least100K US dollars
- Cumulative at some organisations, up to1M US dollars
Source: Google Threat Intelligence Group, 'The Rise of LLMJacking', Nov. 2, 2026
A chain, not a single bug
A Cloud Security Alliance research note dated Aug. 20 reconciles some of this. It cites Obsidian Security's June 11 disclosure of three LiteLLM flaws. CVE-2026-47101 is an authorization bypass on key-management endpoints. CVE-2026-47102 is missing field-level authorization on /user/update. CVE-2026-40217 is the sandbox escape that delivers code execution.
On that reading, a low-privilege internal_user key holder chains the first two to become proxy_admin, and the third turns admin into code execution. That would explain why a standard virtual key could end up with admin rights in Google's case while the advisory says the code-execution endpoint needs an admin credential. The CSA note does not confirm that Google's victim hit this chain, and Google's description of /key/regenerate is not the same as either 47101 or 47102 as the note summarises them.
Which version to run
The CSA note says "Complete fixes for that chain shipped in v1.83.14-stable, released May 2, 2026". The GitLab entry names 1.83.10 as the fix for 40217 alone. Both are earlier than the Nov. 18 listing, so a gateway patched this spring is covered, and one pinned below 1.83.10 has been exposed for months.
Google's own advice is to extend vulnerability management programs to model gateways, "ensuring rapid patching". Spending caps and alerts on max_tokens=1 pings, which it also recommends, limit the bill when a key leaks. A hand-rolled Python sandbox is the weak layer here; Codex's sandbox relies on operating-system mechanisms instead, and our comparison of two labs' cyber tiers covers the access-control side of AI security tooling.
Agencies subject to the catalog have until Dec. 2.
Sources
More in Security
- 01GPGME's CVE-2026-129975 Double Free Needs Only a Crafted Key, and 2.3.1 Is the FixGnuPG disclosed five days early because GitLab had to update, and no CVSS score exists yet to rank it.
- 02Dragos Seeded 2 Million Prompts With Canaries; Amazon's Decoys Caught ARTEX Agents Grading Stolen AI KeysTwo reports from November 13 and 19 show AI credentials being harvested and spent by machines, and a Chinese lab fetching secrets from ZDR prompts.
- 03REDCap's CVE-2026-90817 Needs Only a Public Survey Hash, and CISA Listed It 59 Days After the CVEVanderbilt's survey platform is fixed in 16.0.49 LTS, 17.3.10 LTS and 17.4.4, and Safe Security reports a miner installed with cron persistence and an SSH key.
- 04Ghostscript's Upload2Shell Flaw, CVE-2026-125396, Escapes -dSAFER and Is Fixed in 10.08.0The flaw is scored 7.8 to 8.5 as a local bug, yet researchers confirmed command execution through CUPS print queues, Roundcube message view and WordPress uploads.