The CVE in Google’s LLMJacking Report Points at the Wrong Bug

Mandiant’s incident responders watched a threat actor turn an ordinary LiteLLM virtual key into a full administrative credential, and they named the flaw that allowed it. The identifier they printed, CVE-2026-40217, describes a different vulnerability in the same product, one whose own advisory says it cannot be reached by the kind of user the report describes.

The misattribution does not weaken the incident. It does something more useful: it shows that the bug which actually did the work in these intrusions is the one with no good name, and that the defender reading the report for a patch level will reach for the wrong number.

What the report says happened

Google Threat Intelligence Group published its LLMJacking write-up on 2 November 2026, describing Mandiant engagements across financial services, retail, technology and transportation in North America, South America and Asia Pacific. The money is the part that travels: fraudulent charges “exceeding $100,000 USD in a single day”, with “cumulative charges at some organizations reaching $1 million USD before access was revoked”. One actor made “more than 500,000 LLM API calls in less than eight hours” to power a resale service the report names as jet-night.com.

The escalation step is described in one sentence. A follow-on cluster, “after acquiring pre-validated Virtual Keys from the threat actor that established initial access”, then “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”.

Three facts in that sentence are checkable independently: the endpoint, the mechanism, and the identifier.

The identifier describes something else

CVE-2026-40217, per the NVD record retrieved 3 November 2026, reads: “LiteLLM through 2026-04-08 allows remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI”. Different endpoint, different mechanism. Not a key-management flaw at all.

The GitHub advisory for the same CVE, GHSA-wxxx-gvqv-xp7p, is the one that makes the mismatch load-bearing rather than cosmetic. It confirms the mechanism, “The sandbox can be escaped using bytecode-level techniques, allowing arbitrary code execution in the proxy process”, and then states the precondition in bold: “Reaching the endpoint requires a proxy-admin credential”.

So the CVE cited as the means of becoming proxy_admin is a vulnerability that requires already being proxy_admin. As an account of the escalation step it is circular.

The bug that fits is CVE-2026-47101

The flaw matching the report’s own description of the mechanism has a different number. The NVD text for CVE-2026-47101: “LiteLLM prior to 1.83.14 allows an authenticated internal_user to create API keys with access to routes that their role does not permit”. The GitHub advisory gives the affected range as < 1.83.14, patched in 1.83.14.

Obsidian Security, the research team that found the chain, published the code path. On a key-generate call, “the only thing enforced is that the new key’s user_id matches the caller” — allowed_routes is written to the database unvalidated. And critically for matching the incident, they note the same defect is not confined to one endpoint: “the same unchecked write recurs on every key-write handler” including /key/regenerate, the exact endpoint Mandiant names.

That is the reconciliation. Mandiant’s endpoint is right, Mandiant’s mechanism is right, and the identifier belongs to the code-execution bug further down the same chain.

Why the mix-up is easy, and partly the researchers’ own doing

This is where the subject earns some credit, because the ambiguity is documented rather than invented. Obsidian write that they found a second code-execution path and “We did not request a separate CVE for this variant, and track it under CVE-2026-40217”. One identifier covering more than one code path, in a chain the same team published as a single CVSS 9.9 finding, is a situation in which a responder writing up an engagement months later can reasonably reach for the chain’s number rather than the specific link’s.

It is also worth saying that CVE-2026-40217 is not irrelevant to these intrusions. Once a key reaches admin routes, the guardrail sandbox escape is exactly what is then available, and Mandiant’s own remediation section cites it correctly in that role: “ensuring rapid patching against privilege escalation and authentication bypass flaws (such as CVE-2026-40217)”. The report is internally inconsistent about which bug did which job, not wrong that both were in play.

The practical cost

Patch levels. A defender who reads the incident section, looks up CVE-2026-40217 and upgrades to its fixed version, 1.83.11 per the GitHub advisory, has not fixed the escalation the report describes. CVE-2026-47101 is patched in 1.83.14. Two of the three releases in Obsidian’s account are distinct: 1.83.10-stable carried the role-escalation fix and the replacement sandbox, and 1.83.14-stable carried the remaining allowed_routes fixes. A reader following the report’s identifier lands three patch versions short of the bug that let a standard key become an administrative one.

The second document, and a gap it opens

The detection guidance has its own split worth noting, because it is the half an operator will actually be judged on. For Google Cloud, the report says the relevant inference operations sit in Data Access logs, “which are disabled by default due to volume and ingestion costs”, and that IAM audit policies must be configured manually to capture them. Google’s own Cloud Audit Logs documentation agrees in the same words: “Data Access audit logs are disabled by default”.

Both documents are accurate. The consequence of their agreement is the uncomfortable part: the primary evidence for the abuse stage of an attack that can run up $100,000 in a day is, in the default configuration of the platform being attacked, not recorded. The report’s own framing acknowledges the window, noting that LLMJacking “can generate large spikes in fraudulent cloud charges in a matter of hours”, against a detection control the customer has to switch on first.

What is not checkable here

Several of the report’s load-bearing numbers are single-source by nature and are not independently verifiable: the per-incident charge figures, the 500,000-call count, the eight-hour window, and the attribution of jet-night.com’s pricing to stolen keys, which the report itself frames as a belief (“GTIG believes”) rather than a finding. The victim organisations are unnamed, so none of it can be corroborated from the other side. This article takes those as reported and checks only what has a second document behind it: the CVE records, the vendor advisories, the researchers’ disclosure and Google’s logging docs.

Whether the CVE citation is an error or a deliberate reference to the chain as a whole is also not something the published text settles. Both readings are available. Either way the patch level a reader derives from it is wrong by three releases, and that is the part worth fixing in a reader’s notes rather than adjudicating in the authors’ intent.