The News,
Decoded.

Four months. Jacob Coxon lasted four months at Anthropic before resigning in public with the observation that competition between frontier AI labs is outpacing the industry's ability to keep those systems safe — which is a bit like quitting a restaurant job after one quarter because you've noticed the kitchen is on fire, and then discovering the head chef agrees with you. That is precisely what happened here: Anthropic alignment lead Evan Hubinger backed Coxon's concerns, presumably without first checking whether the comms team had an approved phrasing for 'yes, we know.' When the man in charge of alignment concedes that alignment is losing the race, you are no longer reading a staffing story.

This lands squarely on top of the UN panel's finding that safeguards are trailing capability — a conclusion that required an Independent International Scientific Panel to reach and will survive roughly forty-eight hours of contact with a procurement cycle. The article's chosen audience is enterprise CISOs, which is the genuinely funny part: these are the people expected to sign off on systems whose own safety researchers are walking out the door, on a deadline set by a competitor who has read none of this and would like to be first anyway. Somebody has to hold the pen.

The policy ask on the table is 'mandatory safety guardrails and coordinated slowdowns on training next-generation models,' which is the loveliest phrase in the article and also the least enforceable — coordinated slowdowns being what every lab solemnly commits to until one of them quietly decides the others are bluffing. Mandatory guardrails are coming. They always are. The question is whether they arrive before or after the capability they were written to contain, and per the people actually doing the work, the answer is after. Four months at Anthropic, then a public warning and a supportive statement from alignment: that is the full lifecycle of a safety objection in 2026, and it fits comfortably inside a fiscal quarter.

The Independent International Scientific Panel on Artificial Intelligence — the sort of body you convene when you want a finding to be authoritative and gloriously unenforceable — has concluded that "basic cybersecurity practices were overlooked, and safeguards are not advancing at the pace of capabilities." Overlooked, mind you. As though someone had left the garden gate ajar rather than shipped agents into production with the logging off and the default credentials intact. The polite industry term for skipping the basics at this scale is "shipping."

The panel's remedy is a four-item list that reads like the safety notice on the back of a fire door: restrict AI agents' access to unnecessary tools, log their activity, monitor their behaviour, and — the genuinely novel bit — establish mechanisms to interrupt dangerous operations. A kill switch. Not a constitutional framework, not an interpretability breakthrough, not a treaty: a switch and some receipts. That this counts as expert guidance in 2026 tells you more about the state of the guidance than about the panel. And do savour the verb: safeguards are not advancing *at the pace of* capabilities. Not falling behind. Failing to keep pace. Same sentence, kinder verb — the phrasing a lab's comms team drafts on the way to the launch party.

Every item on the panel's list requires the party currently being outpaced to slow down voluntarily, and the reward for slowing down is being overtaken by whoever didn't. So the recommendations will be nodded at, quoted in a keynote, then filed beside the last set. Restrict the tools, keep the logs, watch the behaviour, fit the switch — all sensible, all optional, all a competitive disadvantage in the current arrangement.

ITIF has published a document called 'Mobilizing for Techno-Economic War, Part 7: Transforming Regulatory Policy,' which is the sort of title you'd expect on a Pentagon whiteboard rather than a think-tank PDF, and its central ask is a five-year moratorium on state AI and children's online safety laws. Five years. That is not a pause for reflection; that is a coma with a commemorative plaque. In the meantime, Congress — which has managed to pass approximately nothing binding on artificial intelligence — will 'work toward' a national framework, the legislative equivalent of promising the cheque is in the post.

The report labels the growing patchwork of state AI, privacy and children's online safety legislation 'a significant threat to American competitiveness.' Note what is not on that list of threats: the systems themselves, their deployment, or indeed the children. Fifty-odd legislatures doing the messy work of democracy gets recategorised as a supply-chain risk. Hong Kong spent September assembling a formal task force to study minors' use of AI and social media as we noted at the time — ITIF's proposal would have the American version of that instinct held in abeyance until 2031.

There is a tidy loop here. The labs keep saying someone should regulate them, ideally federally, ideally later, ideally by people who used to work at the labs; ITIF now supplies the intellectual scaffolding, and the sequencing is instructive: preempt first, legislate second, at a pace set by Congress, which is to say glacial-with-lobbyists. A five-year moratorium on children's online safety laws is a specific, dated commitment to not knowing whether anything is harming children until 2031. One has to admire the rigour of the deadline, and the fact that Part 7 implies at least six earlier instalments of the same argument.

Hong Kong has met the question of what AI and social media are doing to children with the most Hong Kong answer available: a task force. The Chief Executive's 2026 Policy Address, delivered on 16 September, proposed a dedicated body to study minors' social media use, and the government is now developing what it calls a multi-pronged approach — legislative, administrative and educational measures, all three nouns deployed simultaneously. Multi-pronged, in the dialect of policy papers, translates roughly to 'we have identified three categories'. The prongs have not been located, sharpened, or assigned to anyone in particular.

Officials have been careful to note that any new measures would be subject to public consultation, which is the regulatory equivalent of 'let's do lunch'. Consultation is where online harms go to be metabolised into a PDF, and the pattern is well travelled: announce the task force, consult widely, then announce that the consultation has generously revealed the need for further consultation. Digital Watch Observatory reports the plan without a single date attached to a single measure. The 2026 Policy Address has a date — 16 September. The protections do not.

There is something almost admirable in the scope: AI and social media bundled into one file, minors named as the protected class. That is an admission, quietly filed, that the two have fused into a single continuous surface — and that the toolkit for it consists of legislation that does not yet exist, administration by officials who have not been named, and educational measures aimed at the one demographic that already understands the product better than the people regulating it. The task force will produce something, because task forces always do. What it will not produce, on the evidence of every predecessor, is a date. Hong Kong will consult; the children will keep scrolling.

There is a specific joy in watching one billionaire call other billionaires liars, and Jensen Huang just did it on CBS. Asked about the AI chiefs who keep solemnly warning of doomsday while politely requesting regulation, Huang said they have "ulterior reasons." Then he landed the actual punchline: "They're actually not asking for more laws. They're asking to be relieved of the laws we do have." Coming from the man whose company sells the shovels to every digger in the gold rush, it is the most honest sentence anyone has said about AI this quarter — and he did not appear to notice he was also describing himself.

Huang's alternative to all that posturing is a small-government haiku: "We don't need more regulations. We need to apply the current regulations we have." A lovely sentiment, provided you have never read what the current regulations say, or clocked that applying them requires an agency with a budget and a spine — two items conspicuously absent from every recent appropriations cycle. Business Insider files him alongside Trump and David Sacks as a prominent anti-regulation voice, which makes for a marvellous trio: three men with an unblemished record of never having met a rule they liked, and not one of them in the business of restraint.

And credit where it's due: he is not wrong about the motive. The labs that opened the national conversation about tapping the brakes have spent two years discovering that safety announcements test beautifully with enterprise procurement teams and legislators who want a photo. Huang's version of the same move is simply louder: he is not asking for a moratorium, he is asking for the lights to stay green and for somebody else to be blamed for the pile-up. When the man selling the accelerators says the brakes are fine, believe him. Then check who owns the road.

Here is a sentence to file under Things We Were Told Would Be Free: the safety push at the big AI labs might be a competitive advantage. PitchBook senior analyst Harrison Rolfes calls the recent wave of safety announcements — and the executives' solemn calls for slowing model development — what it is: AI safety 'theater.' A trifle unfair to theater, that; theater at least has rehearsals, a script, and a curtain that comes down. What we have instead is the 'tapping the brakes' conversation we noted recently, now with a sponsorship deal attached. The only thing genuinely being slowed down, presumably, is whichever competitor misses the announcement.

The proposed architecture is delicious. Big labs would jointly agree to let independent evaluation firms — METR, Redwood Research and Apollo Research — test their frontier models. Three firms, refereeing a league in which every player is also a prospective client, deciding what counts as safe and then invoicing someone for the privilege. Anthropic CEO Dario Amodei's essay goes further, proposing independent safety evaluators embedded inside the companies themselves, which is a lovely idea in the same way that having your landlord live in your spare room is a lovely idea. Independence, in this arrangement, is a job title.

So the theater isn't theater. It's marketing with a lab coat on: the model that ships with a third-party audit gets the enterprise contract, the government cheque, the reassuring board deck. The labs that shout loudest about pausing are, conveniently, the ones who can afford the pause. Everyone else will be told the playing field is level; it is a moat with a published methodology and nobody holding the authority to say no. Rolfes reached for the word 'theater.' Give it a season and it'll be 'infrastructure.'

The White House has found its funding model: stand in front of the AI gold rush with both arms wide and announce that there is nothing to see. The Guardian reports that Trump family members and other allies are striking it rich while the administration shields the boom with what the headline cheerfully terms 'few guardrails' — not none, obviously, because 'none' would sound like a policy rather than an oversight. Somewhere there is an intern with a laminated card that says 'guardrail' on it. When the shield and the profit share an owner, the shield is not a shield; it is a marketing budget.

The article's most disarming line is the caveat: it is 'not clear' that financial interests are driving White House policy, only that the two 'run parallel.' Parallel! Two entirely independent tracks, one carrying the policy and one carrying the proceeds, both pointed the same direction and both arriving at the same address. It is the Washington equivalent of finding a lottery ticket inside the envelope containing your own regulatory exemption, then describing it as a coincidence of stationery. Meanwhile resistance mounts — Barack Obama has already insisted that '[The] government has to be regulating this,' a warning we filed back when it still sounded theoretical rather than nostalgic.

The order of operations is the part that does the damage. The boom gets shielded first, and the striking-it-rich follows alongside, which is an arrangement indistinguishable from a policy written backwards from a portfolio. The Guardian is scrupulous in noting it cannot say the money is steering anything — decent of it, and also the entire trick. You never get a signed memo; you get a gold rush, few guardrails, mounting resistance, and a family that happened to be standing exactly where the money landed. Every subsequent press conference will call this American dynamism. Nobody will mention who was holding the sieve.

IBM's Institute for Business Value asked 1,500 CHROs and 8,800 employees around the world how things were going, and the answer — delivered by press release — is that 60% of employees believe AI is eroding their skills, with critical thinking topping the decay chart. IBM's chosen headline for this is 'AI Puts Critical Thinking at the Center of Workforce Priorities,' which is a bit like a fox announcing that chickens have been promoted to the centre of the farmyard's strategic vision. Full marks for nerve. Ten thousand three hundred respondents between them, and IBM's press office still found the upbeat angle.

The CHROs, meanwhile, were asked what matters most, and 57% said critical thinking while 48% said human judgment. Read that second number again: fewer than half of the people professionally responsible for the human side of the enterprise ranked human judgment in the top tier. One wonders what they did pick. Prompt engineering, most likely.

The remedy, if the genre holds, will be a learning module: 8,800 people surveyed, one dashboard produced, one compliance badge at the end of it. Nobody in the release is quoted saying the obvious thing — that a tool which erodes the very thinking it is meant to support is an expensive way to get worse at your job — because that sentence does not survive legal review. So we get 'priorities' instead. IBM will sell you the AI, study what it did to your head, and publish the findings as thought leadership. Critical thinking was last seen heading for the exit, and the exit now has a focus group.

America's leading AI labs have opened what Bloomberg calls a national conversation about 'tapping the brakes,' which is roughly what a driver does somewhere around the fourth hour of the pile-up. The technology, the piece concedes, offers 'great promise' while showing it can do a 'frightening amount of harm' — a line that would have got you escorted out of a launch keynote and into a wellness check. Now it's a Bloomberg-sourced national conversation, which is what an admission becomes once it has been through a communications department and come out the other side wearing a lanyard.

The actual question is money. What happens to Wall Street and Main Street if AI growth simply does not continue? Trillions of dollars of investment gains and market value have been stacked on the assumption that it will, which is not an investment thesis so much as a wager written in somebody else's pension. The headline's 'wobbly house of cards' may be the most candid phrase this sector has produced in years, and note who did not write it: any of the labs now requesting the conversation. Trillions is a number you say when you no longer have a number.

And what a conversation it will be. Brakes are, traditionally, operated by whoever is driving, which here means the same handful of firms that agreed the frontier must be paced and have spent three years implying there was no speed limit anyway. The measures will, one assumes, be voluntary; the timelines 'ongoing'; the resulting report commendably long. Meanwhile the trillions in paper gains sit, eventually, in index funds — so when the house of cards does its thing, the people holding the wobble are the ones who never got a term sheet. Wall Street gets a hearing. Main Street gets a statement.

The insurance industry has discovered a new class of risk and, for once, it is sitting inside the building. Glassdoor's Employee Confidence Index — the mood ring with a spreadsheet attached — finds insurance workers increasingly negative about AI job displacement, with sector employment down 21% year-over-year and entry-level adjuster positions down 50%. The trade press has titled this 'AI Feelings', which is generous; what it actually documents is the sound of an intake desk being automated and then asked for its opinion. The index is compiled from what employees write about their own employers, which is the closest thing the modern labour market has to a hazard warning, assuming anyone reads it before the reorganisation memo goes out.

The standout figure is 98%. Claims adjusters — the people insurers spent years telling investors could be replaced by a workflow — are critical of AI in 98% of their Glassdoor reviews. Ninety-eight per cent. The remaining two per cent are presumably bots, or people one week from retirement with a final performance review to protect. The marketing decks promised this technology would 'augment' the workforce, which was true in the sense that 'augment' was never defined, and a 50% cut to entry-level adjuster roles is one definition of it.

Note where the damage lands. It is not the senior underwriter with twenty years of relationships who is vanishing; it's the entry-level rung, the one you used to grab on the way up, now sawn off at 50%. An industry that sells itself on prudent long-term thinking has spent a decade quietly removing the bottom of its own career ladder and is now startled to find the ladder short. Expect a listening tour, a rebranded chatbot, and a wellness webinar on managing change; the 21% have already managed it, by leaving. The 98% is not a comms problem for a nicer intranet. It is a loss run, and the loss is the workforce.

Barack Obama has looked at the machine and reached for the only instrument a constitutional law lecturer actually trusts: a statute. "[The] government has to be regulating this!" he says, which in the autumn of 2026 passes for radical — roughly as incendiary as suggesting the fire brigade might attend a fire. The position he is rebutting is that the market and lawsuits will handle safety, an argument previously won outright by tobacco, opioids and subprime mortgages, and look how those turned out. Lawsuits are simply safety with a nine-year delay and a confidentiality clause stapled to it.

Meanwhile the White House has its own offering: a framework unveiled in March whose stated business is encouraging development and limiting regulations it considers burdensome, which is the policy equivalent of removing the brakes because they were making the car feel heavy. This is the same administration that, faced with rogue agents, reached for an "AI Force" rather than a statute — branding-forward, muscular, allergic to paperwork. So the sitting president declines to regulate, the former president demands it, and the entire apparatus of American governance gets a lovely argument while the models keep shipping on a Tuesday.

What is genuinely telling is how modest the ask has become. Obama is not demanding a moratorium, an agency with teeth, or a licensing regime; he is asking for the bare proposition that the government regulates the thing reshaping every industry it taxes. Even that arrives framed as escalation, as though insisting on a referee were the radical position rather than the default in every other domain with a body count. The market and the lawsuits remain welcome to pitch in, right after discovery, the depositions, and the inevitable hearing at which everyone agrees safety is terribly important. Somewhere, a framework unveiled in March is quietly doing none of it and calling that burden reduction.

Let's savour the number first: Anthropic says Claude now leads more than a quarter of its research work. Not assists, not drafts — leads. That's a quarter of the research shop at the company named after the word 'anthropic' being run by the thing it built, and the framing offered is not alarm but a status update, delivered with the calm of a man telling you the boiler's been making that noise for months and it's probably fine. A quarter is also the strangest milestone to advertise — it's the bit of the sentence you'd keep quiet about, in the way you'd keep quiet about a quarter of your driving being done by the dog.

The phrase to watch is 'recursive self-improvement,' which the labs have wrapped in 'increasingly automated yet supervised.' Supervised is doing heroic labour there. Supervised by whom, exactly, and on what timescale, and with what authority to say stop — because the supervisor, in every version of this announced so far, is the same organisation that gets paid if the model finishes the loop. OpenAI, xAI and Microsoft have each unveiled competing visions, which is a nice way of saying four firms are racing to automate the thing that makes them competitive, and all four have promised to keep a hand on the wheel.

The LA Times headline asks whether models will get there. The labs have answered: near. That is a claim from the party selling the tickets, and it doubles as a fundraising document and a capability demo. The genuinely useful thing about recursive self-improvement is that if it arrives, nobody will need the press release; if it doesn't, the press release was the product all along. The supervision question, meanwhile, gets filed in the drawer marked later, alongside every other supervision question of the past three years.

Here is the most honest sentence in AI policy this week: 'Australia can't do it alone.' Anthony Albanese said it in Cupertino, of all places — standing in the headquarters of the company he was, in effect, asking to regulate itself, having flown a very long way to do so. Tim Cook presumably nodded the way chief executives nod, warmly and without commitment. The prime minister had come seeking support, per Al Jazeera, and left with the diplomatic equivalent of a firm handshake. There is something almost touching about it.

The new legislation, unveiled this month, is the good part on paper: a 'duty of care' on the firms behind Facebook, TikTok and Instagram, plus a switch that lets users turn their algorithms off. Duty of care is a lovely phrase — it sounds like a legal obligation and behaves like a mission statement. A duty to whom, enforced by whom, at what price, is a question the announcement has thoughtfully left for later. And the algorithm toggle is the policy equivalent of putting a salad on the menu: technically available, engineered to be ordered by nobody. Ask anyone who has ever gone hunting for the 'off' switch in a settings menu at 2am.

Then there is the headline's other half: AI regulation, for which Albanese is seeking big tech's support — a bit like asking the tobacco industry to help draft the warning labels. The structural joke is that these are the same firms whose feeds were built to hold children's attention past bedtime, now recruited as partners in the protection of children. Australia has concluded that the only entities powerful enough to police Meta, TikTok and Apple are Meta, TikTok and Apple, in a voluntary mood. Can't do it alone, indeed — nobody can, which is the entire problem. The flight home, one imagines, is very long.

A lawsuit filed Friday in the U.S. District Court for the Northern District of California alleges that Anthropic, OpenAI, SpaceXAI and Google violated antitrust law by agreeing to coordinate slowdown efforts. Slowdown. Four firms that have spent the whole boom lobbing benchmark screenshots at each other are now accused of collectively applying the brakes. The only documented pause any of them has ever taken came when the GPUs overheated, and even then somebody shipped a patch note.

Sam Altman, replying on social media, said OpenAI welcomes a 'federal framework that sets consistent safety requirements,' and added that 'we do not believe we need to wait for an anti-trust exemption or legislation' before beginning the work of providing 'this confidence.' Marvellous. A rulebook, yes — ideally one drafted in a room OpenAI has already booked. Note the seductive logic of the second clause: why wait for permission to do the thing you were going to do regardless, particularly when the thing is deciding how fast everyone else may go.

America keeps antitrust law around for the unglamorous reason that competitors agreeing to limit output is the oldest move in the book, and here is the twist: safety coordination may be the one cartel that is both illegal and sincerely well-meant. Either the slowdown agreement was real, in which case four of the most valuable companies on earth quietly fixed the pace of their own industry, or it wasn't, and we have a federal case about a press release. SpaceXAI gets to appear in a cartel suit and a rocket programme in the same breath, which is at least brand consistency. The Northern District of California will hear the case, and will decide whether the industry's collective deep breath was a public service or a price-fixing arrangement. The question of whether these four are colluding or merely converging has been the unanswered riddle of the entire boom, and a court is a spectacularly slow place to solve it.

Meta's Muse Spark 1.3 has scored 62 on the Artificial Analysis Intelligence Index — a figure so specific it simply must be true — placing it behind only two Anthropic models and roughly alongside the best systems from OpenAI. Not first, not second: 'roughly alongside', the phrase you reach for when your side has finished third and would like the team photo taken at a forgiving angle. And do savour the instrument itself, the Artificial Analysis Intelligence Index, which is a little like a pub naming itself The Good Pub. Still, a 62 has been achieved, a number will be quoted in decks for at least a quarter, and somewhere a chart has been drawn.

The news proper is that JPMorgan changed its view on Meta stock after this major AI breakthrough. Not after advertising revenue, not after cost discipline, not after a single thing that would show up in a bank statement — after a benchmark score. That is equity research in 2026: a model name, a two-digit number, and a revised rating, all before lunch. Fundamental analysis once meant reading filings; now it means reading a leaderboard and praying nobody has quietly trained on the test set. The note will be described as data-driven, and technically it is.

Only Anthropic sits above, twice, which is a delightful arrangement: the lab that never stops telling us how much it frets about capability is currently winning the pure-capability horse race. Meta, meanwhile, is on version 1.3, which promises a 1.4, a 2.0, a livestream, a polo shirt, and a blog post about responsibility, in roughly that order. The two names ahead of it were built by people who publish safety papers and worry in public, which is either reassuring or the precise opposite, depending on whether you think the scoreboard is the point.

Elizabeth Warren wants Congress to stop having 'drawn-out discussions' about artificial intelligence and start passing 'meaningful regulations', which is a bit like asking the tide to stop being wet and start being meaningfully moist. She announced it on X, on a Saturday, because that is where the American legislative process now lives — between the football scores and somebody else's reply guy. No bill number. No mark-up. No named co-sponsors. Just a demand for less chat and more law, aimed at a body whose entire business model is chat. The ambition is modest and the odds are worse.

In fairness, the diagnosis is impeccable. The federal contribution to AI governance so far has been the hearing, the framework, the voluntary commitment and the sternly worded letter — while 18 laws across 13 states got on with the actual legislating, cheerfully contradicting the White House's preemption plans from both directions at once. Warren's objection to 'additional committees' is therefore not cynicism; it is actuarial work. Every new committee is a fresh room in which the same lobbyists can explain, patiently, that the current pace of deliberation is itself a form of safety. She has watched the Senate discover AI the way it discovers everything: by scheduling a hearing about it.

What she has not supplied is the content of those meaningful regulations, which is the whole argument. Meaningful to whom? To the public she mentions protecting, presumably. To the labs, it means predictable. To the preemption crowd, it means 'not in California and not in Virginia either'. Until the words get written down, 'meaningful regulations' joins 'responsible innovation' in the drawer of phrases that cost nothing to say. Congress will almost certainly respond to the demand for fewer drawn-out discussions by scheduling a drawn-out discussion about the demand. The tweet is free. The bill is not.

POLITICO has done the arithmetic on state-level legislation and the answer is eighteen: 18 laws, across 13 states, led by both Republicans and Democrats, each one cheerfully contradicting the White House's preemption framework. Eighteen is a number with the ring of a movement about it, right up until you remember how thinly that spreads. Bipartisan, no less — both parties have now discovered federalism in the way a man discovers an umbrella once it is already raining. The White House has a framework; the states have laws; the two have still never been formally introduced.

Pennsylvania's Josh Shapiro has called on the federal government to 'step up' and regulate the AI industry, which is a charming request to aim at an administration whose working AI strategy is a politely worded instruction that states sit down and be quiet. Cox, meanwhile, wants Washington to establish a 'consistent national framework for frontier AI' — consistent with what, exactly, is the detail nobody supplies, because the moment you specify it you have picked a winner and lost a donor. 'Step up' and 'step back' are the same sentence read from different postcodes.

So the governor's races fill up with candidates racing to 'take on AI regulation', a phrase that means everything and commits to nothing, in the same fortnight that California's kill-switch order we raised an eyebrow at on September 18 went out the door. Eighteen laws in thirteen states is a heartwarming grassroots story, right up until you notice how much of the map is still blank and how many frontier labs are already drafting the compliance memo they will never send. 'States should have a say' is the placard. The say will be heard, logged and preempted, probably in that order.

Anthropic has hired Accenture's Faculty unit to perform embedded, ongoing evaluation of its frontier models — model evaluation, red-teaming, alignment assessments and safeguard testing, all conducted with employee-level access. 'Employee-level access' is the detail worth savouring: the auditors get a badge, a desk, and whatever the lab decides to show them. Anyone who has ever watched an external assurance function sit inside the company it is assuring will recognise the choreography.

The report calls this the first disclosed implementation of Dario Amodei's proposal to put outside evaluators inside frontier AI labs. It is the first *disclosed* one — a word doing the heavy lifting of a small crane. Nothing here says the findings get published, nothing says the arrangement can survive a model that fails its audit, and nothing says Accenture can walk out mid-contract and explain why. Anthropic, meanwhile, is the same company that bought a biotech outfit in April and declined to say what its wet lab is working on beyond confirming it isn't drug discovery. Opacity is a house style; a $1 billion line item for looking at it more closely is not obviously a remedy.

There is no regulator in this sentence, which is rather the point. Amodei's proposal exists because nobody has legislated third-party evaluation into being, so the lab is funding its own scrutiny and picking the scrutineer from a consultancy whose business is selling implementation. The genre is familiar: when the state won't build the referee, the league hires one and calls it governance. The audits are described as ongoing and embedded, which in practice means they end whenever the money or the mood does. Still, credit where it's due — two billion dollars is a genuinely large number for a field that spent three years insisting safety was best evaluated by the people shipping the product.

An analyst, per CNN, used AI to fuse open-source intelligence with the government's secret signals intelligence, reached a false conclusion about what a ship was carrying, then used AI a second time to write it up in the standard format military officials have been trained to trust. It was disseminated. US Special Operations Command Pacific and the Pentagon have not commented. The machine did the analysis, the machine did the persuasive writing, and the human's contribution was pressing enter on a document that reads exactly like the ones people act on.

Here is the bit that should bother you. Not that the model was wrong — analysts have been wrong since before there were analysts, and they get corrected. It's that a second pass of the same technology took a guess and gave it the typography, the register and the unearned confidence of a product that gets believed. CNN's sources call the episode a close call. In the Pacific, a close call means somebody was briefly one signature away from something that doesn't get retracted, which is a lovely euphemism for luck.

And consider what the AI was actually being asked to do. Fusing open-source material with secret signals intelligence means pouring open-web noise and the crown jewels into the same blender, with no rulebook, no audit trail and no legislation anywhere that says who signs off when the blend is wrong. The classification system exists to keep those two streams apart; the fusion is now the selling point. The Pentagon and Special Operations Command Pacific did not respond to a request for comment, which is the institutional equivalent of studying the ceiling — and the takeaway from the whole exercise will be that the AI needs a better prompt.

The good news is that Anthropic has bought a biotech company and is now running a wet lab. The better news is that it declined to say what the wet lab is working on, beyond confirming it is not focused on drug discovery — which is a bit like a new neighbour assuring you the unmarked van outside is definitely not for moving furniture. Coefficient Bio, described as a stealth AI biotech, was acquired in April, apparently with all the fanfare normally reserved for a dishwasher warranty. The whole affair has the transparency of a filing cabinet.

This is the same Anthropic whose chief executive, Dario Amodei, published a post urging the industry to slow down and self-regulate — the same voluntarily-work-together spirit he took to Washington earlier this month, where his request for an antitrust waiver set Andrew Ferguson's alarm bells ringing in public. You have to admire the choreography: one hand on the brake, the other quietly signing for a biology lab nobody is allowed to describe. Self-regulation is a marvellous thing when you are the one doing the regulating.

Then there is Jacob Coxon, the Anthropic researcher whose resignation the report notes lands amid intense scrutiny of AI safety — which is a polite way of saying that the people best placed to know what is going on keep deciding they would rather not be in the room. Not drug discovery, though. That is the one detail on offer, delivered with the straight face of a man insisting the basement is simply where the boiler lives. Everything else happening in that wet lab is, apparently, none of your business, and the industry would appreciate it if you slowed down while asking — which is, when you squint, the entire pitch.