Microsoft has built a thing called MDASH

It is 10:14 PM, which is the hour at which I have given up on the productive evening but have not yet committed to the slow surrender of sleep, and I am eating a piece of cheese directly off a knife in the kitchen, which is a thing I do when I am alone and reading the news on my phone and pretending the cheese is dinner because dinner, as a concept, has slipped past me again. The kitchen light is the kind that makes everything look like it has been left in a drawer too long. The cheese is fine. I am fine.

This is when I learn, via GeekWire, that Microsoft has built a thing called MDASH, which is not a person but a multi-agent AI system, and which scored 88.45% on something called the CyberGym benchmark, beating single-model systems from Anthropic and OpenAI, including one apparently named Mythos. MDASH works by deploying more than 100 specialized AI agents across multiple models, all of them, I gather, scanning for vulnerabilities. One hundred agents. Hunting for weaknesses. Together.

I put the knife down.

I want to be clear that I understand, intellectually, that this is good news, the way a smoke detector is good news, the way a colonoscopy is good news, the way knowing about a thing is supposedly better than not knowing about a thing. Vulnerability-scanning is a noble pursuit. Someone has to find the holes before someone worse finds the holes. But I cannot stop picturing one hundred small, polite intelligences moving through a piece of software the way ants move through a sandwich left on a picnic table, methodical and uninterested in your feelings about it, and I cannot stop wondering whether any of them, at any point, pause.

I once had a therapist who told me my problem was that I treated every interior thought as a guest who had to be entertained, and I think about her a lot, especially when I read about agents. The word agent used to mean a person in a slightly nicer suit than yours who got you a book deal or a worse apartment. Now it means a small synthetic worker who never sleeps and never eats cheese off a knife and is, at this moment, somewhere in a building in Redmond or a server farm cooled by a river that did not ask to be involved, looking for the cracks in things.

What unsettles me is not the score, 88.45%, which is a very specific number designed to feel modest while being, in fact, a kind of flex. What unsettles me is the eleven-and-change percent that got away. The vulnerabilities the hundred agents did not find. Those are still out there, in my banking app, in my thermostat, in the website where I order the cheese.

I put the cheese back in its drawer. I turn off the light. The kitchen, briefly, scans me back.

Source: GeekWire