AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

CISA: attackers are already exploiting a flaw in the LiteLLM proxy

AI-generated contentWritten by AI

cisa.gov, cirt.gov.jm, thehackernews.com, shattered.io, heise.de · 6 sources · Sources published: 18–19 November 2026 · Some source publication dates are unknown

On 18 November, the US cybersecurity agency CISA added four flaws to its catalogue of actively exploited vulnerabilities. One of them affects LiteLLM, which relays calls to language models. Updating closes the flaw but does not show whether the server has already been compromised.

CISA’s catalogue says a remote attacker can use CVE-2026-40217 to run arbitrary code at the LiteLLM URI /guardrails/test_custom_code. As usual, the agency does not disclose the nature or scale of the attacks, heise points out.

LiteLLM is an open-source proxy server. Companies use it to reach language models from many providers through a single OpenAI-compatible API. Shattered.io says organisations keep their providers’ keys, budgets and rate limits in one place on the proxy. A compromised server therefore holds every key it relays.

Administrators use the vulnerable route to test their own guardrails with Python code. A guardrail is a rule that checks content going to or coming from the model. Security firm X41 D-Sec showed that the code can escape the sandbox meant to confine it. Citing Beam’s analysis, Shattered.io reports that the old safeguard searched the source code text for dangerous names such as __import__. The attacker assembled the names only at runtime, so the filter never saw them.

In the default configuration, the attack requires the proxy’s admin credentials, usually the master key, heise and Beam said. The Hacker News cites a report by GTIG, Google’s threat intelligence group, which says attackers used the flaw to elevate limited inference keys into full admin credentials. Their aim was to use model services with other people’s credentials (LLMjacking). Targets included financial and retail companies in North and South America and the Asia-Pacific region.

Severity ratings differ. Beam says GitHub’s advisory rates the flaw moderate, at 5.9 on CVSS 4, while the NVD gives it 8.8 on CVSS 3.1. Shattered.io notes that CISA’s catalogue reflects what attackers are doing and is not a severity rating. Federal agencies must fix or mitigate the LiteLLM flaw by 2 December. The other additions from the same day, Apache Axis, REDCap and Synology DSM, must be fixed sooner, by 21 November.

LiteLLM version 1.83.11 and later contain the fix. Beam and SmartScope say versions from 1.81.8 onwards are vulnerable, but 1.83.10 and later are not. LiteLLM’s official release notes do not mention this cut-off. If updating is delayed, GitHub and X41 recommend blocking the route at a reverse proxy, Shattered.io reports.

The Jamaica Cyber Incident Response Team (JaCIRT) points out that an update does not reveal what happened before it was installed. JaCIRT urges organisations to look for signs of earlier compromise, review the credentials and services the proxy uses, and replace any keys that may have been exposed.

Report an error in this story