Three things happened in AI this week worth slowing down for. One is about hacking, one is about your doctor, and one is about lawyers spending real money. Let's get into it.
GPT-6 Astra Can Find and Build Its Own Hacks
OpenAI dropped GPT-6 Astra this week, and buried in the announcement is a line that should get more attention than it's getting. Astra is the first model to cross what OpenAI calls its "Critical" threshold for cybersecurity. That means it can find vulnerabilities nobody has found yet and build a working exploit for them without a person walking it through each step.
The public version is trained to say no when you ask it to actually write the exploit out. Fine. But the capability is sitting right there now, inside a model with a million token context window that's rolling out to Plus, Pro, Business, and Enterprise users over the next few days.
Here's my take. I reckon guardrails on the public chat window don't mean a whole lot once the underlying skill already exists. Somebody's gonna fine tune around those refusals or just build their own version without them. That's not a reason to panic, but it's a reason to stop calling this stuff a chatbot. It's doing security research work now, and we ought to talk about it like it is.
ChatGPT Just Got Access to 325 Million Patients' Charts
OpenAI hooked ChatGPT Health up to Epic, the system that holds medical records for more than 325 million people. Doctors can now ask it to pull together lab results, medications, old notes, and history instead of clicking through a dozen tabs. It's read only, meaning it can't write anything back into your chart.
OpenAI says it ran the thing past more than 4,300 physician responses across 27 use cases and 99.1 percent came back safe. That sounds real good until you do the math. Out of every thousand answers a doctor gets, nine of them are wrong in a way that matters. In most jobs that's a fine batting average. In medicine that's a real person.
I like that they kept it read only for now. That's the responsible call. My worry is what happens in year two, once doctors get used to trusting the summary and stop pulling the original chart to double check. Nobody keeps double checking a tool forever, especially not one that saves them twenty minutes a patient.
A Law Firm Is Betting $500 Million on Building Its Own AI
Kirkland and Ellis, the highest grossing law firm in the world, is spending around $500 million over the next few years to build a custom AI platform with Palantir. This isn't a subscription to some off the shelf legal AI tool. It's a system built around Kirkland's own case files, workflow, and institutional knowledge, with more than a thousand lawyers expected to use it.
Reports say it takes something like 180 AI engineers and data scientists to pull this off. That's a company sized investment, not a software purchase.
This is the part people miss about where AI adoption is headed. The real advantage isn't which chatbot you subscribe to, it's whose institutional data gets wired into the model first. Kirkland just bought themselves a multi year head start over every firm still waiting on a vendor to figure it out for them. Small and mid size firms without half a billion dollars lying around are going to feel that gap.
That's the roundup for today. A model that can hack, a chatbot that can read your medical chart, and a law firm spending like a tech company. AI is moving fast this week, and I don't think any of us have caught up yet.