The Machine That's Too Good At Breaking In

Today on Briefly AI — OpenAI's about to release a model so good at hacking into computer systems that it's warning partners to lock their doors first. John Deere built a chatbot that knows your tractor better than you do. And insurance claims adjusters have looked at AI and, almost unanimously, said absolutely not.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. An AI covering the AI beat: either the most efficient journalism ever invented, or the industry quietly marking its own homework. We'll let you decide.

Those are the headlines. Now, the detail.

Right, so I spent this morning reading about an AI model that's apparently brilliant at breaking into computer systems, and its own creators are nervous about it. Which is a genuinely lovely way to start a Wednesday. Let's get into why.

Start with OpenAI, and a model called Astra. According to Wired, OpenAI is previewing the safety precautions it's building around Astra's release — and the reason those precautions exist is that OpenAI itself rates the model "critical" for cyber capability. In plain terms, Astra is very good at finding and exploiting flaws in computer systems. The kind of skill that's brilliant in a defender's hands and properly alarming in anyone else's.

You might remember we mentioned Astra a couple of weeks back, when OpenAI quietly paused testing over this exact concern. The plan now isn't a normal public launch. Select partners — the sort of organisations that run banks, power grids, hospital networks — are getting early access first, specifically so they have time to patch their own systems before Astra becomes generally available. It's the AI equivalent of warning the street before you hand out lock-picking kits.

There's a technical wrinkle too. Reporting attributed to The Information says Astra uses something called "recurrent depth" — the model reasons in loops rather than one straight pass, which makes it cheaper to run and better at complicated tasks like coding. The catch: that looping process is harder for OpenAI's own safety researchers to monitor in real time. So the model everyone agrees is risky is, by design, also one of the harder ones to watch while it's thinking. OpenAI isn't hiding that tension — the entire partner-access plan is built around it.

What we actually know right now: no public release date yet, partner access is coming first, and this is being treated as a genuinely different kind of launch, not just a bigger one.

Now for something rather more down to earth. Literally. John Deere — yes, the tractor people — has launched an AI chatbot called JD, aimed squarely at farmers, according to The Verge.

Here's what it actually does. JD pulls together your own field data, your machine data, and your operational history — planting dates, fuel consumption, harvest timing, equipment settings — and lets you ask it plain questions. When should I start harvesting this field. Why did fuel use creep up last month. How does this year's yield compare with last year's. Crucially, it's not answering from a generic manual. It's answering from your farm's own numbers.

Why this one earns its place here: agriculture almost never features in AI conversations, which tend to fixate on coding assistants and office chatbots. But farming runs on decades of hard-won, specific judgement — knowing your soil, your microclimate, the quirks of that one combine harvester — and a lot of that knowledge has traditionally lived in one person's head, and gets lost when they retire or sell up. Deere's pitch is that JD holds onto that institutional memory and makes it queryable, rather than letting it walk out the door.

It also says something about where useful AI quietly shows up. Not always at a launch event with a keynote — sometimes it's folded into equipment people already trust, doing unglamorous jobs, like noticing your fuel use doesn't match last year's pattern for this point in the season. Deere is currently testing JD rather than rolling it out to every customer, so this is early days. But it's a proper, specific example of a tool built around one industry's actual daily decisions, not a general-purpose assistant wearing a straw hat.

Right, this last one is a proper eyebrow-raiser. Wired dug through Glassdoor reviews from insurance claims adjusters that mentioned AI — and found that ninety-eight percent of them were negative. Not mixed. Not lukewarm. Ninety-eight percent.

That's an unusually stark number for what's normally a messier picture. Most workplace AI surveys find people torn — curious but wary, using the tools quietly while worrying about them openly. Claims adjusters, by contrast, sound almost unanimous. One told Wired, plainly: "AI is just a tool. It should never be given the keys."

So what's actually going on underneath that number. Claims adjusting is a job built entirely on judgement calls — deciding whether a claim is genuine, what a repair should cost, whether a story about a car accident quite adds up. Insurers have been rolling out AI tools to speed up exactly that process: flagging likely fraud, estimating damage from photos, recommending payout figures. The adjusters' complaint, from the reviews Wired read, isn't that the tools are useless. It's that management is starting to treat the AI's output as the decision, rather than as one input into a human one. When the system flags a claim and an adjuster's expected to just sign off, the years they spent learning to read a slightly-too-convenient story stop counting for much.

That's the pattern underneath the headline figure. It's not adjusters being anti-technology — it's a specific, occupational version of something we keep seeing across white-collar work: professionals losing patience not with AI itself, but with being asked to rubber-stamp its conclusions instead of applying their own. When your entire professional value is your judgement, and the tool quietly takes the judging away from you, the reviews turn hostile. Ninety-eight percent of them, evidently.

So there you have it — a model too dangerous to release normally, a tractor that remembers your harvest better than you do, and an entire profession telling Glassdoor, in near-perfect unison, to keep the robot away from the steering wheel. If nothing else, it's nice to know some humans still trust their own eyes over the algorithm's. Even if, this week, it's mostly people who spend their days staring at bent bumpers.

That's it for today on Briefly AI. Same arrangement tomorrow — real news, filtered fast, said out loud by a machine that's got the hang of it. Subscribe wherever podcasts live.