decrypted · 19 november 2026 · vulnerabilities and patching · uk policy and law · ai and llm security

A home-made sandbox in an AI gateway made the exploited list

CISA, America's national cyber security agency, added four flaws on Wednesday to its KEV list, the catalogue of flaws known to be exploited in real attacks. One sits in LiteLLM, a widely used gateway that gives staff and applications a single front door to AI models. The fix has existed since April. The gap between fix and exploitation listing is the story.

A guest list that checks names

LiteLLM lets administrators test custom guardrails, small Python checks that screen what goes into and out of a model. The test endpoint ran that Python inside a home-made sandbox. The sandbox worked by reading the code as text and refusing dangerous words. Attackers simply build the forbidden words while the code runs, so the filter never sees them.

Picture a bouncer who checks a guest list by name and waves through anyone who says their name in two halves. The researchers at X41 showed the trick ran commands as root in the default container image. The flaw is CVE-2026-40217, its public reference number. It affects versions from 1.81.8 up to but not including 1.83.10, and the advisory tells operators to upgrade to 1.83.11 or later.

How worried to be

Not wildly, and the sources disagree on the score. The attacker needs a proxy-admin key, so this is not a drive-by. The GitHub advisory rates it 5.9, moderate, on the newer CVSS scale, the zero-to-ten severity score, while one write-up cites 8.8 on the older one. CISA's entry does not say who is exploiting it or how widely.

An AI gateway typically holds the master key, database credentials and the keys to every model provider the firm pays for. A leaked admin key plus this flaw turns a stolen login into control of the machine, and from there into everything the gateway can reach.

What a defender sees and does

The tell is a POST request to /guardrails/test_custom_code carrying strings stitched together from pieces. The quick mitigation is to block that route at your reverse proxy. The proper fix is to upgrade and recreate the container. If you suspect misuse, rotate the master key, database credentials and provider keys, because the process could read all of them.

The Secure by Design lesson

Filtering text to contain code is a losing design, because the attacker controls the text. Real containment means a separate, unprivileged process with no network, or not offering code execution at all. The maintainers swapped the filter for a restriction library, but a separate process is the stronger boundary, in my opinion. Running as root by default widened the blast radius.

For UK directors the practical point is dull. Put AI gateways on the same asset list, patch clock and network segmentation as any server, and keep admin ports off the internet. Seven months from fix to listing is a long time to be one upgrade behind.

Also this week

Synology file servers. Also on Wednesday's list is CVE-2024-10441, a flaw in the software that runs Synology's NAS devices, file servers in a box. Synology's own advisory rates it critical, and its fixed releases date from November 2024, after the Pwn2Own hacking competition. One report ties recent attacks to ransomware, but CISA's entry names no attacker and no scale, so treat that as unconfirmed. The lesson is old: a file server reachable from the internet is a decision, and an unpatched one is a bill.

Apache Axis. A separate entry covers Apache Axis 1.x, an old web services library. Reporting says Apache does not plan a fixed 1.x release and recommends moving to Axis2/Java or applying a source patch, and that Rapid7 saw it used against Confluence Data Center. Libraries like this hide inside suppliers' products, which is the argument for asking vendors for a parts list of what they ship. If a supplier bundles Axis 1.x, ask them now.

Ofcom's child protection duties. On 16 November Ofcom said platforms must ban users who share child sexual abuse material, and must extend hash matching, a database of digital fingerprints of known illegal images, to more services, including pornography sites. Subject to Parliament, the duties are expected from early 2027. Services under the children's codes must either run highly effective age checks or give every user the grooming protections. Ofcom's Oliver Griffiths said firms that do not act "should expect to face serious consequences". Either route means more scanning or more identity data held by someone, so the design question is who holds it, and for how long. Ofcom also dropped planned livestreaming rules, citing the under-16 social media ban due next year.

Sources

If you run AI tooling or file servers and want a second pair of eyes on how they are exposed, get in touch.

More like this

Get the next post by email: subscribe to Decrypted. Double opt-in, unsubscribe any time, or take the RSS feed.

Prefer to listen? Decrypted on Apple Podcasts, or paste the podcast feed into any app.

Play here Loads Apple's player when you press it.