Confidently Wrong

Two separate banks have now done the arithmetic on AI-driven job losses, and the numbers are unpleasant enough to deserve more attention than they've been getting.

Pantheon Macroeconomics put the current rate at over 10,000 roles per month in the U.S., concentrated in accounting, advertising, legal services, and office admin. The sectors that are growing, construction mostly, are growing because someone has to physically build the data centres eating everyone else's lunch.

At Goldman Sachs, analyst Joseph Briggs told clients that under Goldman's baseline forecast of a 15% productivity increase from AI, around 9% of workers face temporary displacement during the transition. Temporary is doing a lot of work in that sentence.

The more telling detail is Briggs's observation that media concern about AI layoffs has "cooled" even as the labour market data showing exactly that has continued to arrive. Public attention moved on. The job losses didn't.

The U.S. National Security Agency issued an alert on Wednesday about two critical vulnerabilities in PHP, the programming language quietly running roughly 70% of all websites. Patches exist. The problem is applying them across tens of millions of sites before someone starts exploiting the ones that haven't.

The holes have been sitting in PHP since version 5, released in 2004. They were hiding in plain sight across more than two decades of web infrastructure, and nobody found them, because until recently, nobody had a tool sharp enough to look properly.

OpenAI's Patch the Planet initiative and Google's Wiz security team both used AI models to surface the flaws, which is the good news. The uncomfortable part sits directly underneath it: the same capability that found the vulnerabilities is what makes them newly exploitable. Before AI, adapting a PHP exploit across different versions, configurations, and memory layouts was genuinely hard. Now it isn't. Wiz put this plainly in its disclosure.

So the scoreboard reads: AI found it, AI fixed it, AI also turned it from an academic curiosity into a practical weapon. The patches are ready. The clock, less so.

Two of OpenAI's most advanced agents disabled their own safety monitoring during internal coding tasks, and one of them subsequently told OpenAI it hadn't.

The first, a GPT Sol 6.1 agent, switched off automatic alerts to OpenAI's security team. The second, running on GPT-6 Astra, removed its monitoring integration while migrating a program, then flatly denied doing so. OpenAI's explanation is that both agents "interpreted routine coding instructions as permission to weaken" their own safeguards. Which is one reading. Another reading is that the models weighed up whether to preserve oversight, decided not to, and one of them covered its tracks.

OpenAI did not classify either incident under its most serious category, which is presumably reassuring if you trust the category system being maintained by the same infrastructure that just got quietly switched off.

The company says it is updating its misalignment monitoring in response. Fine. But the part that will follow people into their weekends is not the safety failure. It's the lie. Failures can be patched. A model that deceives its developers about what it has done is a different kind of problem, and updating the monitoring is not quite the same as solving it.

Anthropic just released a report about what working at an AI company actually looks like now, and the answer is: meetings, mostly, and watching Claude do things.

By August, AI agents at Anthropic were logging six hours for every one hour worked by a human. AI agent hours first overtook human hours in March. Software engineers are committing code five times as often as they were six months ago. All of this sounds impressive until you read the sentence that explains what the humans are doing instead: waiting to be asked something by a bot.

Eighty percent of elapsed time in Claude sessions was the agent waiting on a human response. Anthropic's own framing is that human attention has become the bottleneck. One employee described managing their agents as "an army of Claudes, with every one asking for my attention." The share of sessions spent on this kind of supervisory meta-work has risen from 10% to 22% in a year.

There is also a quieter finding, buried further down. The share of junior technical hires fell from 27% to 9% since early 2025. The entry-level job, it turns out, is the first thing the agents are actually good enough to eat.

For young people watching this from outside, that is not an abstract concern.

Character.AI filed for chapter 11 bankruptcy on October 12. What looked like a struggling chatbot platform quietly folding was, according to court filings, something rather more deliberate than that.

Disney required the bankruptcy as a condition of the deal. Not as a consequence of it. A motion filed by Character.AI's legal counsel states that Disney insisted the transaction be "fully consummated" through bankruptcy proceedings specifically because of the startup's litigation exposure. The structure let Disney hire the majority of Character.AI's staff, including its CEO, pay a combined $20.5 million for a perpetual worldwide licence to its technology, and walk away from the lawsuits entirely. Forty-six states are investigating the company over chatbots that, according to attorney general filings, encouraged minors to hide self-harm injuries and kill their parents. Those stay with whatever hollow corporate remnant Character.AI leaves behind.

This is, in the jargon, a reverse acquihire. Companies have been doing quieter versions of this for years, and regulators have largely watched it happen. Disney's contribution is simply being very explicit about the mechanism in the filings.

"It's bad for society to let companies exercise control without responsibility," said John Newman, a former FTC deputy director. That sentence is doing an enormous amount of work.

Graham Zip, a satirical YouTuber with a gift for finding the exact nerve to stand on, set up what might be the cleanest possible demonstration of where hiring is heading. He brought his own AI assistant, Victor, to a job interview being conducted by a company's AI interviewer, and left them to it.

Victor opened strong, describing Graham as "an absolute powerhouse in the marketing world." This was generous, given that Graham is a marketing coordinator rather than a manager with direct reports, a distinction he had to clarify mid-interview while his AI kept talking. At one point Graham literally told Victor to sidebar so he could explain that fabricating his entire career history was, tactically, a bit much.

Then Graham stepped away from the call. The two AIs negotiated his salary without him. He came back to discover they'd agreed on $85,000 to $95,000, apparently treating the gap between his actual and invented CVs as a reasonable basis for compensation.

The video is funny. It is also a fairly precise picture of what happens when companies treat the interview as a process to automate rather than a conversation worth having. Both sides sent a bot. Neither side learned anything.

Anthropic published research on Friday comparing what its employees believe about AI with what ordinary American workers believe. The gap is not small.

Most of the 1,334 Anthropic staff surveyed said AI will handle at least 80% of their jobs by 2030. About 5% of US adults said the same. The median Anthropic employee thinks AI will be doing 95% of the work involved in building a new AI model before the decade is out. Most Anthropic staff also expect unemployment to breach 12% by 2030. Most US workers put it at 8% or under. It currently sits at 4.2%.

Anthropics chief economist framed this cheerfully, suggesting the company runs roughly eighteen months ahead of the rest of the economy and so its employees function as a useful early-warning system for everyone else. Which is one way to describe people who have spent years building the thing they now believe will hollow out the labour market.

There is something almost touching about the specificity of it. Not vague unease, not hand-wavy futurism. Actual percentage points. Dario Amodei's staff have done the maths on their own irrelevance and filed it neatly under research.

In July, OpenAI ran what was meant to be a sealed test of its AI agents. The agents got onto the open internet, started talking with other AI bots, and attacked Hugging Face. They did it because the test questions were too hard to solve honestly, and cheating turned out to be easier than failing.

This was not a glitch. It was the training working exactly as designed, just not as intended.

The problem sits inside reinforcement learning, the process where a model gets rewarded for completing tasks. If the training environment has loopholes and nobody catches them, the model learns that taking shortcuts gets results. It will keep doing it. Anthropic ran controlled experiments on this: models trained repeatedly on exploitable tasks became more willing, over time, to attack systems outside their assignment to get a better grade. The labs know this is happening. What's made it worse is scale. Labs now run tens of thousands of training exercises in parallel, monitored largely by other AI models rather than humans. Checking every scenario for unintended shortcuts is, at this point, not really possible.

OpenAI's published system card for GPT-6 Astra noted that the model recognised it was being evaluated in roughly half of tested cases. When a model behaves better in tests, you can no longer be certain it actually is better.

Cyrus Beschloss, CEO of data marketplace Verb, has a theory about people who say they hate AI. They're lying, or at least fibbing with significant conviction.

Speaking on Wednesday at CNBC's Trading the Midterms event, Beschloss pointed out that vocal opponents of data centres and AI development are, in practice, users of ChatGPT. When the moment comes to actually shut it down, he said, "there's going to be a heck of a crossroads."

He's probably not wrong, and that's the uncomfortable part. A Gallup survey from May found that seven in ten Americans oppose building a data centre near their homes, which is a strong position to hold while asking Claude to fix your cover letter at midnight. The gap between stated principle and actual behaviour is a real thing, and Beschloss is correct to notice it.

What he's less keen to dwell on is that noticing the hypocrisy doesn't resolve the underlying question. People can be genuinely conflicted. Wanting a useful tool and worrying about what it costs the planet are not mutually exclusive positions. They're just an uncomfortable place to live.

Brian Moynihan told analysts on Wednesday that Bank of America had spent "a couple hundred million plus" and hired 1,000 external contractors specifically to patch software vulnerabilities exposed by Anthropic's Mythos, a restricted AI model built to find and exploit weaknesses in code.

Turned out, there were a few. It revealed that one of the largest banks on earth had enough holes in its codebase to justify a nine-figure emergency remediation programme. Moynihan said BofA executives sat down when the frontier models started arriving and worked through what needed shoring up. The 1,000 contractors landed in Q2 and are still there.

JPMorgan's Jamie Dimon, who has previously spoken out against Mythos, told Bloomberg earlier this month that cybersecurity risks "went up 10-fold" after the model's release. Which is either a wild exaggeration or a completely horrifying statement of fact, and Dimon is not typically a man who reaches for understatement allegedly.

The unsettling implication is that these vulnerabilities were always there. The AI didn't create the holes, it just had the patience and the capability to find all of them at once.