A dependency you ship is being exploited. What now?

When a dependency lands on CISA's known-exploited list, check whether you run it, then patch or block it. A short check for startups.

Discuss a project

By Keith Rull ·

Find out whether you run it, then upgrade or block the exposed path, and look for signs it was already used. CISA’s known-exploited list means the bug is in use somewhere, not that every named product is in your stack. Finish that check this week and leave the rest of the catalog for later.

The entry that matters for a small team shipping model calls is LiteLLM. Here is what the public record says, and the check I would run.

What changed this week?

On 18 November 2026, CISA added four bugs to its Known Exploited Vulnerabilities catalog. The agency says it adds a bug when it has evidence of active exploitation. One of the four is CVE-2026-40217, which CISA names as a BerriAI LiteLLM unprotected alternate channel. The same alert also names Synology DiskStation Manager, Apache Axis, and Vanderbilt’s REDCap. The notice is CISA’s 18 November alert.

The day before, CISA added one more: CVE-2026-115731, a Perforce Puppet Enterprise bug. That notice is CISA’s 17 November alert. I am not walking through how either bug is used. The useful fact is that both are on the list because someone is already exploiting them.

CISA’s notice points at Binding Operational Directive 26-04, a rule for federal civilian agencies. The notice says that rule applies to those agencies, and that CISA encourages every organization to fix catalog bugs first. A startup is not on the federal clock. The catalog is still the right short list, because it means the bug is in use, not merely that a description was published.

Most of the names will not be yours. A startup that does not run Synology, REDCap, Axis, or Puppet can note them and move on. The one I would not skip without a search is LiteLLM. Small teams put it in front of model calls, often in a container they have not opened since the demo.

What should you check first?

Start with the product name, not a write-up of the bug.

Search your repositories, lockfiles, and container images for litellm. If the name is absent, write down the date and the places you looked. Do the same for the other names on the two alerts if there is any chance you run them. A miss is a result. It is how you avoid a week spent patching software you do not have.

If litellm is present, read the vendor advisory before you pick a version. GitHub’s advisory GHSA-wxxx-gvqv-xp7p, retrieved 20 November 2026, is the record the catalog cites. It was published on 11 May 2026. It says the POST /guardrails/test_custom_code endpoint runs Python a user supplies, inside a sandbox that can be escaped, so code can run in the proxy process. It says the default Docker image runs that process as root. It also says that, in the default configuration, reaching the endpoint requires a proxy-admin credential.

The same record does not agree with itself on two points. Do not smooth them over.

The score and the privilege sentence disagree. GitHub lists a CVSS 3.1 score of 8.8, and that vector says an attacker needs little privilege. The impact section says an admin credential is required in the default setup. Both lines are on the advisory. I would still patch. A stolen admin key is enough to remove the control the impact section is counting on.

The version lines disagree too. The affected range is litellm from 1.81.8 up to, but not including, 1.83.10. The record names 1.83.10 as the first patched version, and a link points at the v1.83.10-stable release. The patch paragraph says the fix is in 1.83.11 and tells you to upgrade to 1.83.11 or later. Meet the stricter line and stay outside the vulnerable range: install 1.83.11 or later. Read the advisory again before you ship. A vendor can edit the page.

If you cannot upgrade today, the advisory’s workaround is to block POST /guardrails/test_custom_code at your reverse proxy or API gateway. Blocking the path is a mitigation. It is not evidence that the bug was never used.

This table is a way to sort the finding. It is not an audit of anyone’s system.

What you find Do this Do not do this
You do not run the product Record the name and the date you looked Open a project to patch the whole catalog
You run a version below 1.83.10 and can upgrade Install 1.83.11 or later and keep the previous image Stop at 1.83.10 because one line calls it patched
You cannot upgrade today Block the custom-code endpoint and limit the admin key Leave the proxy open until the next sprint
You import the library and do not run the proxy Confirm the process is not listening, and write that down Assume an import is the same thing as the proxy

How do you decide what to fix?

Use exposure, not the length of the catalog.

If a stranger on the internet can reach the proxy, treat the upgrade as this week’s work. If the proxy is only inside a private network, still treat it as this week when the admin key is shared, old, or sitting in a repository. If the key is narrow and the port is private, schedule the upgrade and block the endpoint now. If you do not run the product, stop.

The admin key is a production credential. Every AI agent needs a name and an end date is the rule I want on that key: a named owner, a narrow scope, and an expiry. A proxy that can run code is a bad place for a key nobody can retire.

Keep the previous image so a bad upgrade can be turned off. How to keep an architecture decision reversible is the note for that path: the behavior you can restore, who can restore it, and what a rollback will not undo. The same habit is already on a practical security baseline, in the dependency row and the practiced rollback. This week’s check is that baseline pointed at one named bug.

If the proxy is how an AI feature reaches a model, the patch is part of the feature, not a side errand. The feature still needs a fallback when the proxy is down for the upgrade. Take the proxy offline only after that fallback has been tried.

This example is hypothetical. A small startup proxies chat through LiteLLM in Docker, pinned at 1.83.3, with the admin key in an environment variable. The port is open on the private network and closed to the public internet. That version is inside the vulnerable range. The same day they block the custom-code path, rotate the admin key, and put 1.83.11 on a branch that can roll back to the previous image. They write down the date, the old version, and who looked. They leave Synology alone. Nothing in that sketch is a client incident.

If you already know which image runs the proxy, I would treat the search and the block as an afternoon, and the upgrade as another day when you build the image yourselves. If a vendor image wraps the proxy, the upgrade waits on that vendor and the block should not. That timing is an estimate, not a measured engagement and not a price.

What would I do?

I would search for litellm before I read a description of the bug. A missing name gets a dated note, and I move on. A hit gets 1.83.11 or later, or a block on the endpoint the same day if the upgrade cannot ship. The new admin key gets an owner and an end date. I redeploy the previous image once in a test environment so the rollback is a memory.

I would not buy a scanner to answer this. A scanner helps after you know which image is in production. It does not tell you whether the package is there.

You can do this without help when you can name the image and the key. Ask for help when the proxy has been reachable from the internet, the admin key is shared, and nobody can say who last used it.

FAQ

Does the federal directive apply to a startup?

No. CISA’s 18 November notice says the directive applies to federal civilian agencies. The same notice encourages everyone else to fix catalog entries ahead of other bugs. You are not on their deadline. The reason to move is that the bug is already being used, not that a federal date has arrived.

Is version 1.83.10 safe?

The advisory does not give one answer. The version range treats 1.83.10 as the first patched release. The patch paragraph says to upgrade to 1.83.11 or later. Install 1.83.11 or later so both lines are satisfied, and re-read the advisory before you ship. A vendor can correct a page after you have read it.

What if we import the package and do not run the proxy?

The impact the advisory describes is the proxy endpoint, not a function call in ordinary application code. If you are not sure the proxy process is absent, treat it as present until you can show it is not listening. Write down where you looked.

Should we patch every new catalog entry?

No. Check the product name against what you actually run. If you do not run it, record that and stop. The catalog is a priority list of bugs attackers are using. It is not a backlog you have to finish in a week.

If a dependency you ship is on the exploited list and you cannot tell whether you run it, talk to us.

Build your startup’s MVP, and be able to run it

Discuss a project