Five holes entered the actively exploited catalogue in two days

I opened the catalogue and checked all five numbers one by one before writing anything. They are all in there, exactly as the feed gives them.

The CISA KEV catalogue is not a list of discovered holes. Hundreds of those come out every week. Only what has confirmed use in real attacks goes in here. The difference between the two is the difference between could happen and already happening.

Which of the five I am watching

Ray. The other four matter, but they are the usual suspects: operating system, virtualisation, documents, network service. Every serious admin knows them by heart.

Ray is different. It spreads computation across machines, and that is exactly why it sits under a large share of model training and serving. Code injection there means somebody else's code on the machines holding your weights, your data and your keys.

AI infrastructure is now widespread enough to have its own holes in the actively exploited catalogue.

That is the quiet signal in this whole story. Two years ago a catalogue like this took in routers, mail servers and office suites. Today it takes in a framework for distributed machine learning, because there are enough installations to make it worth an attacker's time.

What the deadline means

The 20 and 21 August dates are an obligation for US federal agencies. For you they are not an obligation. But they are a good measure of how fast somebody who has seen the real attacks thinks you should move.

Three days. That is what you would have been given in their shoes, and that number says more than any severity score.

If you run vCenter or SharePoint, you know what to do. If you run Ray and it never occurred to you that it is part of your perimeter, now is the moment for it to occur to you.

The visual is generated code-art, no third-party images.