CISA added four actively exploited flaws to its Known Exploited Vulnerabilities catalog yesterday, November 18. Three of them have a federal deadline of this Friday, November 21. The fourth is the one I want to talk about first, because it's an AI gateway.
The Four
| CVE | Product | What it is | Federal deadline | |---|---|---|---| | CVE-2026-40217 | BerriAI LiteLLM proxy | Code execution through the custom guardrail testing endpoint (CVSS 8.8) | Dec 2 | | CVE-2026-90436 | Apache Axis 1.x | Unauthenticated RCE via unsafe reflection (CVSS 9.8), no fix coming | Nov 21 | | CVE-2026-90817 | Vanderbilt REDCap | Code injection via public survey routing and data import (CVSS 9.8) | Nov 21 | | CVE-2024-10441 | Synology DSM | Unauthenticated RCE (CVSS 9.8), used in C3RB3R ransomware attacks | Nov 21 |
LiteLLM: The Keys to Every Model
If you haven't run into LiteLLM, it's a popular open-source proxy that sits between your apps and the model providers. Your teams talk to LiteLLM, and LiteLLM holds the real API keys for the cloud models behind it. That's what makes it convenient, and it's what makes it a target.
Per The Hacker News, citing a Google Threat Intelligence Group report, attackers have used CVE-2026-40217 to turn limited-access inference keys into full administrative credentials and to keep access, as part of LLMjacking campaigns. The flaw is in the guardrail testing endpoint (/guardrails/test_custom_code), which can be abused to run code on the server. Targets include financial services, retail, transportation and technology organizations across North America, South America and Asia Pacific.
LLMjacking is a boring crime, which is why it works. Google and Mandiant described it earlier this month: attackers steal API keys or cloud credentials, then resell the model access through "LLM mixer" services. The victim gets the bill as if it were legitimate usage. In some cases attackers also used victims' cloud accounts to host their own proxies and workloads.
This is a practical problem, not a sci-fi one. An AI gateway is a credential vault with an HTTP interface. Treat it like one.
Apache Axis: The Library Nobody Knew They Had
Axis 1 is end-of-life, and Apache doesn't plan a fixed 1.x release. This is the same Axis bug that powers the 4EverAble2Shell chain against N-able N-central I wrote about this week, where it's been used to drop cryptominers. GBHackers, citing Rapid7, says it's also been exploited in Atlassian Confluence Data Center where the legacy XML-RPC and SOAP remote APIs are enabled; Atlassian Cloud isn't affected. The fix there is upgrading Confluence to 9.2.28 or an unaffected 10.x release.
The real question is where else Axis 1 is hiding. It's 20-year-old Java plumbing, and it shows up inside commercial products you'd never think to check.
REDCap and Synology
- REDCap is research data capture software used heavily by universities and hospitals. Exploitation requires a valid public survey hash, which, for a public survey, is often exactly what's public. Safe Security reported an unknown actor using it in late October to drop a web shell and an XMRig miner disguised as syslog-ng, with cron persistence and an added SSH key. Unaffected versions per GBHackers: 16.0.49, 17.3.10 and 17.4.4.
- Synology DSM is the old one, a 2024 bug that I covered earlier this month when Synology tied it to C3RB3R ransomware. The KEV listing just makes it official.
What To Do
- LiteLLM: don't expose the proxy or its admin endpoints to the internet. Upgrade to the current release and check the project's advisory for the fix. Rotate every provider key it holds and every virtual key it issued, review admin users, and set spend and rate limits with alerts at the provider. Google's advice applies broadly: deny model access permissions that aren't explicitly needed, use a secrets manager, and watch inference and audit logs for spikes or attempts to turn logging off.
- Apache Axis: search your environment and your vendors' products for
axis.jar1.x. Migrate where you can; where you can't, cut network exposure to the SOAP endpoints. Confluence Data Center admins: patch, and turn off the legacy remote APIs if you don't use them. - REDCap: upgrade to a fixed version, then hunt for web shells, unexpected cron jobs, new SSH keys and processes pretending to be syslog-ng.
- Synology: update DSM (and BeeStation OS) and take the login page off the internet.
I work at Zscaler, so the usual bias: I'd rather none of these be reachable from the internet at all, and for most of them there's no business reason they need to be. Patching still comes first.
Bottom Line
Three of these are classic "old thing on the internet" stories. The fourth is new: the box that holds your AI keys is now a known, exploited target. Put it behind access controls, cap the spend, and watch the bill like it's a security log, because it is one.
Sources: The Hacker News, GBHackers, BleepingComputer on Google's LLMjacking research, CISA alert