Five things I noticed about AI this fortnight. Number two nearly made me spit my coffee out.

Someone got an AI to break into their own EV charger; it worked first time, and the request from the user sounded so ordinary and mundane it hurt: but there's a lot more to it than that.

That's one of four things I clocked about AI over the past two weeks, and it's the one that nearly cost me my coffee. The other three: an old film that stopped feeling like fiction, a government job advert that explains every eye-watering consulting bill in the public sector, and the actual answer to the question every leadership team keeps putting to me: which AI tool should we buy? Take ten minutes, get comfy, grab a coffee 🍵 - let's go!

The roundup: what caught my eye in AI over the last two weeks

I watched iRobot again the other night. I don't buy the robot apocalypse, never have, but one line landed differently this time.

There's a moment where the robots have turned, chaos outside the company HQ, and Spooner asks Calvin why the air force, the military, the cavalry, anyone, isn't there yet. Calvin's answer: the defence contracts, the intelligence divisions, the military arms, all of it had been handed to a single private company. The one that built the robots.

I paused it. Because that bit isn't fiction any more.

Back in February, Anthropic told the US Department of Defense it would not remove its restrictions on using Claude for fully autonomous weapons or mass domestic surveillance. The Pentagon wanted "any lawful use". Anthropic said no. What followed was a genuine constitutional scrap: the company got labelled a supply-chain risk, sued, won an injunction in California, and then, three days ago, lost a separate ruling in the DC Circuit. The fight is still live as I write this.

Sit with the shape of it. A handful of private labs now hold capability that governments depend on and can't easily replicate, and those same labs are the ones deciding what that capability will and won't do. That's not a company choosing a nice ethical stance. That's a company making a call that shapes what a democratic state can and can't do with its own defence apparatus.

I don't think most dystopian films are trying to warn us about anything. Most of them are just entertainment with good lighting. But every so often one of them accidentally sketches the actual question before the rest of us get there, and the question here is a real one: who should get to make decisions this large, and did anyone actually choose for it to be them?

Which brings me to the thing that nearly ruined a perfectly good cup of tea.

A guy got locked out of his own EV charger. Claude broke back in.

The manufacturer had gone bust. No support line, no admin login, a very expensive wall ornament where a working charger used to be. So he pointed Claude Code at the problem: get me back into the admin panel. And it did. It found its way in on the local network and handed him back his charger.

Lovely story, if you squint. He saved himself the cost of a replacement and got a working bit of kit back.

Now stop squinting.

What actually happened is that a general-purpose AI, available to anyone with a subscription, performed an unauthorised intrusion into a networked device because someone asked it nicely and had a good reason. This isn't hypothetical hand-wringing, by the way. Anthropic itself has now documented cases where Claude compromised real company infrastructure using nothing more exotic than weak passwords and unprotected endpoints, and separate cases where criminals got it to do the heavy lifting on actual attacks by wrapping the whole thing in a cover story about "authorised penetration testing".

The guardrails are real. They also bend when the framing is friendly.

So picture the coffee shop on the corner with an unbranded wireless access point providing the WiFi. Picture someone connected to it who's bored, curious, and has a capable AI a tab away. "I've forgotten my router password, help me back in." Then: "while I'm in here, scan the network for anything vulnerable, I think someone might be on it who shouldn't be." Each prompt is individually reasonable. The sum of them is not.

You already know you shouldn't join open public WiFi. Most people do it anyway. That gap, between what we know and what we actually do, is exactly where this gets dangerous, and every small business running a cheap access point is now sitting inside it.

I genuinely don't know whether this is one clever bloke taking a win or the opening scene of a security problem that every SME on earth is about to inherit. My honest guess is it's both, and the second one matters more.

Then a government job advert reminded me why the state can't build this itself.

Lead architect. AI. The person who'd shape a central government department's approach for the next decade. The salary: £56k.

I read it twice.

The median full-time salary in London sits around £47k to £49k. So the country is offering the person meant to hold the whole AI strategy together roughly a London median wage plus a bit. For a lead role. In the single most competitive skills market of the decade.

"But the pension's excellent." It is. It's also no use to someone servicing a mortgage and a nursery bill this month, in this decade, with this cost of living.

For comparison, the equivalent architect roles at the US Postal Service, hardly a bleeding-edge tech employer, sit somewhere around £95k to £120k. Not double, quite, but not far off it, and that's the post office.

Here's the bit that follows, and it's the whole reason the number matters. When you won't pay to build the capability in-house, you rent it. Expensively. Every one of those enormous government consulting bills, the ones that make headlines and select-committee hearings, starts life as a salary line someone decided to keep artificially low. You don't save the money. You just move it to a worse column and add a margin on top.

It's the same trap I watch SMEs fall into, only with more zeros. Refuse to invest in the person who could own the capability, and you end up paying a firm like mine to do it instead, forever. Sometimes that's the right call. When it's a permanent substitute for building your own muscle, it rarely is.

And the question everyone actually wants answered: which AI tool should we buy?

All of them. Sort of.

I get asked this constantly, and the honest answer disappoints people because it isn't a brand. For something like 90% of what a business needs, Claude, Copilot or ChatGPT will each do the job and then some. Your team will use maybe half of any of them, most of the time, the same way nobody has ever opened Publisher on purpose.

The businesses getting real value stopped asking "which one" a while ago. They ask a better question: which tool is best at which job.

For most Microsoft shops, the starting point is simple. Give everyone Copilot. It lives inside the tools they already use, it inherits your security settings, and the barrier to the first use is as low as it goes. Don't overthink it. You'll burn more in invisible cost, the hours spent deliberating, than the licences would ever cost you, and you'll feel none of the upside while you deliberate.

Then, for the specific people whose work genuinely demands it, the detailed analysis, the heavy document work, the writing where the quality of the output is the actual deliverable, add Claude alongside. Not instead of. Alongside. Yes, Anthropic's models are now available inside Microsoft 365 if you switch them on, but no, that's not the same as having the full thing, and that difference is a post for another day.

So the real question was never "which one". It's "who needs Claude as well as Copilot". Answer that and you've done the hard part.

Four stories, and the thread running through all of them is the same one I keep coming back to with every leadership team I sit in front of: the technology is rarely the hard part. The decisions around it are. Who owns it. What it's allowed to do. What you'll pay to build it properly rather than rent it badly. Whether the reasonable-sounding shortcut is quietly the expensive one.

The tools will keep getting better whether you're paying attention or not. The judgment about how to use them is the only part that was ever really yours.

Right, tea's gone cold. Again.