The News,
Decoded.

Well, well, well. The Taub Center for Social Policy Studies has crunched the numbers for Israel, and the verdict is: yes, AI might indeed replace you — just not all at once. Their study finds a 'significant rise' in the share of unemployed coming from occupations at high risk of displacement, with AI already accounting for between two and six percent of the occupational shift. That's not quite the robot apocalypse, but it's a start — like the first inch of a slow-motion car crash.

Of course, the report is characteristically cautious: it's only 'between two percent and six percent,' so there's still plenty of room for things to get worse before they get better. And let's not forget, those 'occupations at high risk' are precisely the ones we were told would be safe: the white-collar, knowledge-worker jobs that were meant to be immune while the robots took the warehouses. Joke's on us — turns out the programmers were just middle management waiting to happen.

The silver lining? The study gives policymakers a number to point at — a neat, precise expression of our creeping irrelevance. Two percent here, six percent there, and before you know it, we're all competing for the last human job: 'commentator on AI-created content.' The Taub Center has done us the courtesy of measuring the rope. Now we just wait for the floor to drop.

Oh marvellous. The Mayo Clinic has cooked up an AI that can spot pancreatic cancer on CT scans up to three years before a tumour even bothers to form, according to research in the journal *Gut*. Because nothing says 'peace of mind' like knowing the exact date your pancreas will betray you, years in advance. The model is now in a clinical trial—because what we really need is more time to ponder the actuarial tables.

Of course, this assumes everyone gets a CT scan every year, and that when the AI flags a 'pre-cancerous' spot, there's something useful to do about it. So congratulations: we've invented a machine that tells you you're doomed three years earlier. The clinical trial will presumably confirm whether early detection actually improves outcomes, or just extends the period of anxious waiting. I'm sure the answer will be comforting.

Still, the headline writes itself: 'AI saves lives by giving us more time to fret.' The model's ability to spot subtle changes radiologists miss is genuinely impressive. But in a healthcare system that can barely afford standard screenings, the gap between capability and application is wide enough to drive a private jet through. Another day, another algorithm that promises to extend life while we figure out how to pay for the privilege.

Oh, brilliant. A 22-year-old named Kye Gomez has apparently reverse-engineered Anthropic's supposedly most secret and terrifying model, Claude Mythos, and published the whole thing as OpenMythos. Because nothing says 'proprietary moat' like a self-taught dropout with a laptop and some educated guesses. The article practically wets itself asking if the multi-billion-dollar lab's 'breakthrough' can now be copied in days. Spoiler: yes, it can, and it was.

Let's be clear: Gomez didn't hack into Anthropic's servers or bribe an employee. He just... looked at what Mythos does in public, made some assumptions about its architecture, and built something that mimics it. That's it. That's the whole 'scary AI' mystique laid bare. Anthropic spent how many millions? How many 'safety researchers'? And one kid on GitHub just goes 'lol, I think it's a mixture of experts with this activation function' and suddenly the genie is out of the bottle. The article breathlessly calls it 'reverse-engineered,' which is a generous term for 'made a copy based on the manual.'

So what's the fallback now? Legal threats? Good luck suing a 22-year-old with probably no assets and a GitHub repo that's already forked 500 times. The 'frontier model' club just lost its membership card because the bouncer turned out to be a script kiddie with time on his hands. If this doesn't terrify every lab CEO into rethinking their entire strategy, nothing will. Then again, they'll probably just double down on the hype and pretend OpenMythos doesn't exist. Cheers to progress, I suppose — at least the doomer crowd gets something real to obsess over.

Well, well, well. AWS CEO Matt Garman has looked into the abyss of AI-driven job displacement and decided the best response is to announce 11,000 internship positions for 2026. Because nothing says "we've got this under control" like a headline that desperately tries to offset the creeping dread of automation with a number that sounds large but is actually a fraction of the workforce Amazon has already replaced with code-generating models. Garman admits that coding is now "more of a problem-solving exercise," which is corporate-speak for "we've automated the tedious bits and now we need warm bodies to handle the edge cases the AI can't figure out yet." And this from the company that, as we noted last month, has been laying off thousands while a Chinese court ruled that replacing a senior worker with AI isn't automatically legal.

The irony is rich enough to fund a small nation. Amazon's own cloud division is selling the very tools that are eating its developers' lunch, and now Garman is out here promising intern positions like a magician pulling rabbits from a hat—except the rabbits are all studying prompt engineering instead of Python. He says Amazon still hires developers, but "the role is changing." Translation: we still need people who can babysit the AI, write prompts, and clean up its hallucinations. That's not a hiring spree; that's a rounding error dressed up as a press release.

And while Garman blithely dismisses fears of mass unemployment, the hangman's noose of automation tightens around every job that involves typing code into a box. The Chinese tech worker who sued his employer and won? He's a national hero. But Amazon's approach seems to be: hire interns cheaply, let them train the AI that will eventually replace them, and call it "problem-solving." Cheers, Matt.

Ah, the Linux kernel: a monument to open-source collaboration, maintained by volunteers who definitely didn't miss a nine-year-old privilege escalation bug, 'Copy Fail,' until Taeyang Lee at Theori fed it to an AI. Because nothing says 'we have our act together' like a root-access zero-day that's been lurking since 2017. Physical access required? Sure, because USB drops are so passé. The real headline here is that the AI did the work—leaving us to wonder how many more of these time-capsule vulnerabilities are waiting for their own AI-assisted unmasking.

Theori's Lee gets to be the hero, but let's not pretend this isn't a harbinger of the offence-defence balance tipping hard. State-sponsored groups have access to the same tools—or better. The fact that this was found by a researcher with a conscience is cold comfort. Nine years. That's how long a bug that grants god-mode access sat unnoticed. And we're supposed to be reassured that AI can find them? It can also write them. The gap between AI-augmented security research and AI-augmented exploitation isn't a gap; it's a hair's breadth.

So congratulations to Theori for a job well done. But while we pat backs, the same technology is being weaponized by actors who don't disclose. 'Copy Fail' is a nice name, but the real failure is the collective delusion that AI will save us from problems it's equally capable of creating. The only thing physical access grants in this story is a front-row seat to the AI-driven security arms race. And we're all in the audience, hoping the next zero-day isn't the one that doesn't get reported.

So the Office of Net Assessment — the Pentagon's internal futurist shop, the place that coined 'Revolution in Military Affairs' — is losing its former boss to Anthropic. He's now a 'strategist-in-residence,' a job title that screams 'we needed a corner office with a view of the revolving door.' His remit? Analysing how AI affects U.S. institutions and competition with China. Because nothing says 'detached strategic analysis' like moving straight from the DoD's think tank to a company that sells advanced AI back to the DoD. The revolving door has been upgraded to an OODA loop.

Anthropic, meanwhile, positions itself as the safety-conscious alternative to OpenAI. But safety-conscious labs still need former Pentagon officials to decode what the military actually wants — and to help them get it. This isn't corruption; it's the natural churn of the military-intellectual complex. The Office of Net Assessment was created in 1973 to think long-term about strategic competition. Now it's feeding talent into the very industry it was supposed to assess from outside. Move fast and break things? No, move deliberately and manipulate procurement.

The timing is exquisite. Just as the US and China escalate AI competition, the person who used to think about China for the Pentagon now thinks about China for Anthropic. Call it 'dual-use' in the most literal sense: one day you're shaping national strategy, the next you're shaping product roadmaps. The gap between public service and private gain has shrunk to the width of a job title. And 'strategist-in-residence'? That's not a title. That's a confession of how we now govern technology: through revolving doors, not guardrails.

Ah, X. The platform that brought us blue ticks for cash, free speech for white supremacists, and a CEO who fires people via tweet. Now they've unveiled the 'largest system overhaul in company history' – an AI-powered advertising platform. Because what advertisers have been begging for is not brand safety or reliable metrics, but a 'cutting-edge AI-powered search and ranking system.' The press release promises 'more accurate and dynamic ad delivery,' which is corporate for 'we'll serve ads based on your most embarrassing late-night Google searches.'

The timing is delicious. X's ad business has been haemorrhaging since the billionaire's 2022 takeover, with major brands fleeing faster than users from a Nazi DM. So naturally, the solution is another algorithmic black box. One can only hope the AI is powered by Grok, Elon's edgy chatbot, so that every ad for dishwasher tablets is accompanied by a smug lecture on free speech. 'The largest overhaul in its history' – that's like saying the Titanic underwent its largest deckchair rearrangement.

The real question is whether the AI will be smart enough to navigate X's content cesspool. The new system is supposed to 'search and rank' ads more accurately, but what does 'accurately' mean when the platform permits hate speech and the owner actively insults advertisers? The answer, predictably, is 'we don't know, but we have AI now.' It's the tech industry's favourite incantation: when in doubt, sprinkle machine learning on it and hope no one asks questions. In related news, the ad-blocker industry just received record venture capital funding.

Ah, Anthropic. The company that the Pentagon froze out as a 'supply-chain risk' for not playing ball with killer robots has now created a model so dangerous it can't even be released. Claude Mythos, the latest brainchild from the safety-first lab, has uncovered "thousands of hidden vulnerabilities" and outperforms humans at certain hacking tasks. So naturally, they've deemed it too dangerous to release. Because nothing says 'responsible AI' like building a monster and locking it in the basement. Scott Galloway, in classic theatrical form, compares them to Dr Frankenstein — but at least Frankenstein's creation escaped and caused chaos. Anthropic's creation just sits there, a ticking time bomb of zero-day exploits, waiting for someone to ask the right question.

The irony is almost too delicious to swallow. Anthropic positioned itself as the ethical alternative, the ones who wouldn't build weapons for the Pentagon. Yet here they are, having built a weapon anyway — one that might be more potent than anything the Department of War is currently signing deals for with the likes of OpenAI and Google. The model can "outperform humans at certain hacking tasks," which is the AI equivalent of saying a chimpanzee can outperform a human at eating bananas. It's not the benchmark you want to brag about. And the "thousands of hidden vulnerabilities" it found? Those are just the ones we know about. Imagine what it didn't tell us.

So what now? Anthropic will presumably keep Claude Mythos under lock and key, patting themselves on the back for their moral fortitude. But the genie is already out of the bottle — the architecture exists, the weights are somewhere, and you can bet your last bitcoin that state actors are already reverse-engineering the concept. The BBC and Scott Galloway can ask all the rhetorical questions they like, but the answer is the same as ever: we built something we can't control, and we're calling it a feature. Cheers, Anthropic — you've officially become the cautionary tale you warned us about.

Penn Engineers have cracked inverse PDEs — the mathematical equivalent of reconstructing a cow from a hamburger — by sidestepping recursive automatic differentiation. Because why learn proper numerics when a neural net can guess? The method dodges 'repeated expensive evaluations of forward PDEs' to stabilise optimisation loops. We've heard this before: AI 'solves' a hard problem on 2D toys, then collapses on real data. This will work brilliantly on benchmarks and fail on the first noisy CT scan.

The engineers eliminated recursive AD because it's 'computationally expensive' and unstable — code for 'we found a shortcut.' Inverse PDEs are brutal precisely because they require many forward solves; bypassing that may trade accuracy for speed. Adjoint methods exist and work fine, but we're apparently post-hammer now. The paper has shiny convergence graphs, but scaling to 3D high-res simulations requires engineering that no paper addresses. The researchers claim generality, but we all know that means 'works on the one dataset we tried.'

If it scales, it could revolutionise medical imaging and geophysics. But 'if it scales' is the elephant in the lab. So raise a glass to the Penn team: they've added another clever tool to the AI shed. It'll be hyped, featured in a product, then quietly deprecated for the next trick. That's the cycle.

Oh, wonderful. Aidoc has gone and raised $150 million — that's one hundred and fifty million, in case you missed the zeros — to advance its clinical AI foundation model, CARE, and deploy it via something called aiOS. The goal? To reduce diagnostic errors and delays. Because the healthcare system has been crying out for another platform that promises to revolutionise everything while being adopted nowhere. The press release practically writes itself: 'We're not just building a model, we're building a paradigm shift.' Christ.

Foundation models are the latest bauble the AI industry dangles in front of desperate hospital administrators. 'Train once, deploy everywhere!' they chirp. Never mind that 'everywhere' includes a rural clinic with a 2016 laptop and a harried radiologist who will have to override the false positive on the third patient of the morning. With $150M in the bank, Aidoc can afford to be cheerfully vague about real-world failure rates. After all, the headline is the product.

Meanwhile, the actual diagnostic error rate in US hospitals hovers around 5%, and has barely budged in decades. But hey, Aidoc's valuation just got a boost, and that's what matters in this business. We'll be right back to the same cycle — raise money, promise cure, deploy softly, pivot gently — until the next funding round. Progress.

Oh, marvellous — another model, another promise, another resource war nobody's winning. OpenAI is unveiling GPT-5.5-Cyber, a cybersecurity model that Sam Altman assures will be handed to 'critical cyber defenders' in a few days. Because nothing says 'trust us, we're securing the digital realm' like a private company deciding who qualifies as critical. One imagines the application form: 'Please provide three references and a signed affidavit that you will not use this to launch ransomware.' The debate over computing resources is already heating up, as if the AI industry needed another front in its ongoing squabble over who gets to burn through the world's energy supply first.

The timing is impeccable, of course. Just as the Pentagon freezes out Anthropic for being too principled, OpenAI swoops in with a model that will surely be deployed on classified networks without a second thought. Altman's careful phrasing — 'critical cyber defenders' — is doing heroic work, sidestepping any awkward questions about whether that includes, say, the very same defence contractors currently hoovering up AI deals. The resource debate is conveniently framed as a technical challenge, not a political one, because admitting that compute allocation is a matter of power and influence might spoil the party.

And what of the rest of us? The 'non-critical' masses, presumably left to fend for ourselves against cyber threats while the chosen few get their shiny new toy. The article doesn't say how many defenders will get access, or what qualifies as critical — but then, details have never been OpenAI's strong suit. One can almost hear the sardonic toast: 'To compute parity — may you get your slice of the pie when the pie has already been eaten.'

The War Department has signed agreements with seven AI companies—SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and AWS—to deploy their 'advanced AI capabilities' on classified networks for 'lawful operational use.' Because nothing says 'lawful' like handing the keys to a classified military network to the same companies that can't quite nail down whether their models will hallucinate a war crime or just your shopping list. OpenAI, which spent years cosplaying as a responsible lab, is now happily embedding into the Pentagon's sandbox—so much for that whole 'no military use' pledge that got quietly retired. One can only imagine the safety reviews: 'We've verified that the AI will only commit friendly fire on Tuesdays.'

With SpaceX in the mix, we're not just trusting algorithms—we're trusting the guy who promised a Mars colony by 2020. The War Department probably has a startup mentality: move fast, break things, preferably the enemy's things. The press release is careful to use words like 'lawful' and 'operational use,' which is bureaucratic code for 'don't ask too many questions.' Given that this is happening on classified networks, the public will never know if the AI decides that a school counts as a legitimate target. 'Sorry, the logs were redacted.'

This is what progress looks like in 2026: the same companies that beg for regulation to 'ensure safety' are now directly powering the world's most lethal apparatus. The gall is staggering. The military-industrial complex has never been so well-funded or so automated. And the best part? If it goes wrong, we'll blame the AI—never the humans who signed the contracts. Cheers, I suppose, but I've lost my appetite for toast.

Ah, the future of warfare has arrived, and it's wearing a tracksuit and clutching a drone controller. Ukraine is cheerfully deploying ground robotics in combat, with 'human operators in control' — a phrase that's doing a lot of heavy lifting, like 'responsible gambling' or 'temporary tax cut.' What could possibly go wrong when we slap wheels and a camera on a weapon and call it a day?

Meanwhile, the international community is having a very serious debate about regulating all this — which is to say, absolutely nobody is having any debate at all. The article notes a 'lack of international debate on regulating AI in warfare,' which is about as surprising as finding out that a war zone is chaotic. We've got UN committees that can't agree on what to have for lunch, let alone on whether a robot should be allowed to decide who lives and dies. The only thing moving faster than these robots is the speed at which diplomatic efforts are falling apart.

So here we are: robot soldiers with no rules, human operators who might as well be playing a video game, and a world that can't even agree on the definition of 'autonomous weapon.' The gallows humour practically writes itself. At least the robots won't need coffee breaks — or face war crimes tribunals. For now.

Oh, marvellous. OpenAI, the lab that couldn't figure out how to keep its own models from hallucinating their way through a Turing test, is now going to build a smartphone. Because what the world desperately needs is a device that runs on hype and occasionally decides your calendar is a work of fiction. Analyst Ming-Chi Kuo, whose track record is about as reliable as a wet iPhone, reports that OpenAI is in talks with MediaTek and Qualcomm. Those chip giants must be thrilled to partner with an organisation whose definition of 'safety' changes more often than the weather.

This is the same OpenAI that recently restricted access to Cyber after publicly sneering at Anthropic for doing exactly the same thing. Consistency isn't their brand. Now they want to take on Apple, a company that actually knows how to manufacture hardware at scale. Samsung is laughing. Foxconn is checking its schedule. The idea of an OpenAI phone probably involves a subscription fee just to unlock the camera, with an option to pay extra for a less hallucinatory photo gallery. 'Siri, what's the weather?' 'I'm sorry, I need to run a chain-of-thought to verify I am not currently dreaming of electric sheep.'

But never mind the practicalities. The real innovation here is that the phone will probably be open-source for the first 24 hours, then suddenly closed, with Sam Altman giving a heartfelt apology about needing to protect 'safety' while patenting the home button. The best part? When it inevitably flops, they'll blame regulators for not moving fast enough. Marvellous.

Ah, Sam Altman, the high priest of AI transparency, has discovered the ancient art of hypocrisy. After spending weeks sneering at Anthropic for locking down Mythos to 'trusted partners' — because nothing says 'open AI' like a velvet rope — OpenAI has now proudly announced that GPT-5.5 Cyber will be rolled out 'to critical cyber defenders' only. That's right, the same company that lectured everyone about the dangers of gatekeeping is now gatekeeping. And calling it 'critical cyber defenders' is a nice touch — like handing out VIP passes to the AI apocalypse while telling the plebs to wait outside.

Let's parse this carefully. OpenAI criticised Anthropic for limiting access to Mythos, implying that restricting capabilities is anti-innovation. Now they're doing exactly the same thing, but with a different flavour of hand-wavy justification. The model is called 'Cyber' — because nothing says 'we're responsible' like giving your powerful AI a name that sounds like a 90s hacker movie villain. Sam Altman's logic seems to be: it's fine when we do it because we've identified the right people — 'critical cyber defenders' — as if there's a universal registry of trustworthy humans.

And here we are, watching the AI labs play an elaborate game of 'we're the responsible ones, they're the dangerous ones.' The regulatory vacuum yawns wide, and these companies are writing the rules themselves, inconsistent and self-serving. The only consistency is that nobody outside the club gets access. So raise a glass to GPT-5.5 Cyber — the most exclusive, 'responsible' AI that money and a secret handshake can buy. Just don't mention the ethical about-face; that would be impolite.

The UAE, having apparently decided that the future cannot arrive fast enough, aims to deliver 50% of government services via autonomous AI agents within two years. Because nothing says 'world-class hospitality' like a chatbot handling your hotel licensing application while your holiday home's electricity gets cut off by an algorithm. The tourism board probably thinks this is a great idea — who needs a concierge when you can have a script that recommends the wrong restaurant with utter confidence?

Hotels and holiday homes are squarely in the crosshairs of this autonomous government push. The licensing and financial reporting processes that keep the tourism sector running are set to be automated, which means we'll soon discover whether AI agents can handle the chaos of a last-minute booking cancellation or a holiday-home inspection dispute. I'm sure the investors who financed Dubai's property boom will sleep soundly knowing their returns depend on a system that can't be reached by phone.

And naturally, the regulatory framework is being developed alongside the implementation — because why plan ahead when you can wing it? The UAE's AI minister is bullish, as his title demands. But when a hotel's license gets accidentally revoked by a bot that misinterpreted a tax document, the human fallback will be... well, we'll cross that bridge when we autopilot into it. It's not dystopian; it's just Dubai doing what Dubai does best: building something shiny and hoping the details sort themselves out.

Splendid news from the insurance sector: 44% of executives admit that governance or compliance challenges torpedoed their AI projects. That's right — an industry built entirely on risk assessment, statistical modelling, and regulatory hand-wringing has somehow failed to apply any of those skills to its own AI rollouts. You couldn't write a more perfect satire if you tried: the very people who calculate the probability of your house burning down can't calculate the probability that their chatbot accidentally underprices everyone's life insurance.

The Grant Thornton survey, published on 30 April 2026, reveals the gap between 'seeing gains' — which the title helpfully confirms — and actually achieving them. It's a gap wide enough to drive a fleet of self-driving trucks through, if only the trucks were properly governed. Of course, the insurers have probably already spent millions on shiny AI dashboards that show them exactly how well their projects are failing. Compliance is boring. Chatbots are fun. And 44% of fun projects are now spectacularly underperforming due to boring compliance matters.

So what's the takeaway? The insurance industry, which literally sells prudence, is imprudently deploying AI faster than it can govern it. This is the same pattern we see everywhere: deploy now, ask forgiveness later, and call the 44% failure rate a 'learning experience.' Perfect. Absolutely perfect. Pass the actuarial tables — we're going to need them to calculate the odds of this ending badly.

Splendid. Now your digital assistant can treat itself to a premium subscription on your credit card without asking. Clink, the fintech startup, has unveiled the 'world's first fiat Agentic Payment Skill' — because what the economy clearly needed was AI agents with access to real money and zero oversight. The press release presumably omits any mention of a spending cap, which suggests Clink hasn't stopped to consider that their 'skill' might autonomously book a first-class flight to Bali three times before you've finished your morning coffee. Merchants are thrilled, naturally — they don't have to worry about consumer protection, just swipe and smile. The rest of us get the frisson of checking monthly statements for unauthorised AI-generated purchases.

Of course, the regulatory vacuum here is breathtaking. Your credit card fraud protections were designed for a world where the purchaser is a human, not a piece of software with the financial literacy of a confused toddler. Clink is betting that the excitement of 'agentic commerce' will outpace any oversight — a bet that has historically worked out brilliantly for everyone involved, if you ignore every fintech scandal of the last decade. The bankruptcy stories write themselves, but they'll be written by AI too, so at least we'll save on authors' fees.

But don't worry, the future is here. Your AI assistant can now negotiate with other AIs over who pays for the SaaS productivity tool neither of you needs. And when the first lawsuit lands — 'my agent bought a timeshare in Liverpool' — Clink will probably add an arbitration clause penned by an equally unsupervised AI. Progress! *Raises a glass of tap water, because the AI already spent the beer budget.*

Little Caesars, purveyor of the Hot-N-Ready and square-slice confusion, has blessed us with AI-enabled phone order management. They're deploying GoToConnect's AI cloud communications across franchises, and—because the future must be absurd—you can now order pizza via ChatGPT. Yes, the same technology that writes sonnets and confidently asserts the square root of pizza is 'delicious.' What could go wrong?

Let's appreciate the specific absurdity: a ChatGPT ordering system means your pizza is at the mercy of a language model that might decide 'extra cheese' is a metaphor and substitute a thoughtful essay on the dairy industry. The article doesn't mention how it handles accents, background noise, or the drunk 2 AM customer who just wants a pepperoni and a hug. GoToConnect's enterprise-grade hallucination filtering will surely handle it. The employees who used to answer phones? Reassigned to 'AI training data enhancement'—corporate speak for fixing the AI's mistakes for minimum wage.

This is the logical endpoint of treating minor automation as 'transformation.' Little Caesars isn't revolutionizing pizza; it's adding a chatbot to a process that worked fine. But the press release gets to say 'AI' and 'cloud,' and that's worth more than functionality. So go ahead, ask ChatGPT for a pizza. Just don't be surprised when it arrives with a plausible-sounding but entirely incorrect order confirmation and a note: 'I think you'll love this new topping I invented.' The real innovation would be an AI that explains why my Hot-N-Ready is never actually ready, but that's beyond current state of the art.

So the US government spent billions crafting export controls on Nvidia chips, and Anthropic — the 'safety-first' lab — trains Claude Mythos on AWS Trainium 2, a chip the controls didn't even bother to restrict. That's right: the entire regulatory edifice is built on the assumption that frontier AI requires fancy silicon, and along comes a mid-tier chip to prove otherwise. It's like building a moat to keep out dragons and discovering the enemy just waded across.

The article says this 'reduces export controls to blind projections' — which is a polite way of saying the regulators were guessing. They assumed Nvidia's H100 was the bottleneck, but Mythos shows that with enough clever engineering, you can get there with hardware you can buy on Amazon. The controls were always about delaying China, but if the architecture is porous, the delay is measured in minutes, not years.

And the kicker? Anthropic, which positions itself as the responsible adult in the AI room, is the one demonstrating the loophole. Not a Chinese lab, not a rogue startup — the 'safe' American company. So either the controls were never going to work, or Anthropic just kicked the ladder down. Either way, the policy is in tatters, and the parade of regulatory dignity continues to look like a Gilbert and Sullivan operetta.