Latest

Cybersecurity

LiteLLM and LLMjacking, the AI software that grants full access

LiteLLM enters the CISA catalog over a flaw that turns an inference key into administrative access to the LLM.

Translated from Spanish by an automated system. The Spanish original prevails. Read the Spanish original

LiteLLM and LLMjacking, the AI software that grants full access
LiteLLM and LLMjacking, the AI software that grants full access

An API key for a language model should be good for asking questions and paying the bill. In LiteLLM, the open source proxy that many companies place in front of OpenAI, Anthropic or Google to spread spending across teams, that same key can end up granting administrator access to the proxy itself. CVE-2026-40217, with a severity of 8.8 out of 10, entered the CISA catalog of exploited vulnerabilitieson November 19, which gives U.S. federal agencies until December 2 to fix it.

The detail that makes it relevant is not the score, lower than the other three in the same batch, but the compromised asset. Whoever controls LiteLLM controls the keys of all connected providers, the record of who spends what and the policies that decide which model handles each request. Google's threat intelligence group places the exploitation in LLMjacking campaigns, a term that describes the theft of inference capacity to resell it or use it without paying, with targets in banking, retail and technology.

The technique is not exotic either. The Bytechap chronicle on the CISA batch describes the abuse at the `/guardrails/test_custom_code` endpoint, a utility for testing validations before enabling them. There lies the lesson for any team deploying AI gateways.

LiteLLM and why its key is worth so much

LiteLLM solves a practical problem: giving each team a single address to talk to, with a common format, and deciding behind the scenes which provider each request goes to. Along the way it stores the real keys, applies spending limits per user, logs consumption and can intercept content with its guardrails. In practice, it is the cashier of a mid-sized company's AI budget.

That is why the impact of a flaw there is measured in three layers. The first is money, because a stolen inference key is resold or consumed until someone looks at the bill. The second is data, because requests pass through that proxy and its log retains what was asked. The third is trust, because an attacker who reaches the panel can add their own model or disable a policy without the end user noticing anything odd in the response.

A nearby precedent is worth keeping in mind. In August, the LiteLLM 1.107.0 package stole keys on Python startup when it was published compromised: the supply chain and the server configuration are two doors of the same building.

The flaw lives in a testing URL

The description of the flaw is short: the `/guardrails/test_custom_code` route allows rewriting bytecode through an unprotected URI. Bytecode is the intermediate code executed by the Java or Python interpreter, and whoever can replace it can change what the program does without touching the original file. The documented result is a privilege escalation that takes a limited inference key to administrative control of the service.

That leap separates an annoying incident from a serious problem. A key with a spending limit should not be able to create users or read other people's logs, and with this vulnerability it can. The prize is not the key found, but the ones behind it.

The December 2 deadline set by CISA is longer than that of the other three vulnerabilities in the batch, and the difference makes sense. Axis, REDCap and Synology received until November 21 because their exploitation is tied to ongoing mining and ransomware campaigns, while LiteLLM is associated with the silent abuse of computing capacity.

CVE Product Severity CISA deadline Observed use
CVE-2026-40217 BerriAI LiteLLM 8,8 December 2 LLMjacking and escalation to administrator
CVE-2026-90436 Apache Axis 1.x 9,8 November 21 Cryptomining on Confluence
CVE-2026-90817 Vanderbilt REDCap 9,8 November 21 Web shells and disguised Monero miner
CVE-2024-10441 Synology DSM 9,8 November 21 C3RB3R ransomware on NAS

Four flaws and the same economic pattern

What is interesting about the November 19 batch is that the four vulnerabilities serve the same economy. Apache Axis 1.x is end-of-life and has no official patch; Rapid7 documented its use against Atlassian Confluence Data Center installations with the remote API enabled, and the attacker did not stop at executing commands: they downloaded scripts from temporary tunnel domains, enumerated the environment, entered MySQL with Confluence's own credentials, searched for card numbers using the Luhn check and tried to harvest credentials with Kerberos.

In REDCap, Vanderbilt University's survey management tool, the path requires a valid public survey hash, a requirement that sounds like a barrier and in practice is obtained from any open questionnaire. Safe Security documented transient web shells and Monero miners disguised as image files, the kind of camouflage that survives a superficial review of the server. In Synology, the 2024 flaw is used to deploy C3RB3R ransomware on network storage devices, which are precisely the ones nobody patches because they have been running for years without complaining.

Four products and an uncomfortable conclusion: no attack starts with a brilliant mind, but with a forgotten piece of the inventory. The N-able N-central case and its unauthenticated shell showed the same mechanics in managed providers' software.

What such an attack really seeks

On an inference server, the target is rarely the server. The attacker wants the computing capacity and the credentials that pay for it, and for that it is enough to maintain discreet and stable access. Spinning up a miner or encrypting a disk draws attention; adding a model with a name similar to a legitimate one and diverting a percentage of traffic does not.

That is why the useful response is not just updating. First it is worth knowing who can reach the proxy's testing route from the internet and whether the keys it manages have spending limits or are provider master keys.

December and the weight still to come

December 2 closes the CISA deadline and, with luck, most installations will have updated. Two unknowns remain. The first is how many companies have exposed the proxy to the internet unnecessarily, a pattern we already saw with the remote management software that N-able left without authentication and that repeats every time an internal tool is published on a panel so that it is "easier."

The sector has precedents to look at itself in the mirror. The analysis of the 948 flaws OpenAI found in Linux left the same question on the table, and the answer was not technological but organizational.

The other unknown is whether the sector accepts that its gateways are de facto critical infrastructure. Treating them as such requires inventory, per-user logging, key rotation and configuration review before every deployment, the same treatment an exposed database receives.

Until that happens, the AI key someone copied into an environment variable will remain the cheapest door in the building, and the inference log will be the only witness when the bill arrives.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
Bitcoin, the Fed and Nvidia: the keys to the crypto week

Keep reading