The News,
Decoded.

Missouri has decided to legislate what should be obvious: that a chatbot cannot marry you, own property, or serve as your CEO. The bill, which would ban government entities from treating AI as a conscious person with legal personhood, is a triumph of something – though probably not clarity, given that it had to be written at all. One imagines the legislative session, with lawmakers gravely debating whether a large language model can file for divorce. The fact that this is necessary says more about our collective grip on reality than any AI capability: we've become so used to treating algorithms as 'intelligent' that we need a law to remind ourselves they don't have souls, spouses, or stock options.

This is the regulatory gap as performance art. While frontier labs push toward AGI, state legislatures are busy clarifying that you can't marry a GPT-7 or appoint it CEO – which, honestly, is a relief for the actual people who'd be out of a job otherwise. The bill's sponsors probably think they're being proactive, but they're essentially passing a law that says 'water is wet.' The private sector, meanwhile, is free to grant AI whatever weird rights it likes – and we know some startups are already trying to bootstrap 'AI co-founders' into equity agreements. So Missouri's ban applies only to government entities, meaning the madness can flourish in the private sector.

Which brings us to the real joke: the bill is a reaction to a problem that barely exists, while the actual regulatory challenges – training data disclosure, liability for harm, election interference – are left to the same slow grind. Missouri's lawmakers might as well ban school buses from flying to Mars. It's a distraction from the fact that no one in power is seriously grappling with the near-term risks of AI. But hey, at least we've settled the great 'can my copilot be my spouse?' debate. Cheers, Missouri – you've solved everything, again.

Well, here's a plot twist that'll give tech executives a nice alibi. Researchers have published a study claiming work-from-home exposure is a 'stronger predictor of weaker junior hiring' than generative AI exposure. How convenient. Just as public anxiety about AI job displacement peaks, along comes a paper that lets CEOs off the hook: 'Don't look at our AI-powered chatbots, it's the remote workers eating entry-level jobs!' It's the perfect corporate deflection — blame the WFH staff for not mentoring, not the algorithm that now does half the junior analyst tasks.

The study, covered by Business Insider, pinpoints the decline in entry-level roles more on remote work challenges than on AI. So the message is: no need for pesky AI regulation, just force everyone back to the office and junior hiring will magically recover. Of course, this conveniently ignores that many companies are actively replacing junior roles with AI tools — the researchers just happened to find a stronger statistical signal from WFH. But statistics are funny like that; they can show correlation while missing the elephant in the server room.

So what's the real takeaway? Tech firms can now say they're protecting jobs from AI while demanding return-to-office orders — a two-for-one special in dismissive management. And the entry-level workers? They'll be crushed between the AI that replaced their application screening and the senior staff who refuse to train them over Zoom. Cheers to progress, where the biggest threat to a junior job is the inability to overhear office gossip. But at least it's not the robots, right? Right?

Well, here we go again. Security firm X41 D-Sec has discovered a critical vulnerability in Starlette, an open source framework that underpins millions of AI agents—because what could possibly go wrong when your autonomous software agents rest on a free, community-maintained web stack? The flaw allows hackers to breach servers and access sensitive data, which is fine, really, if your AI agents aren't doing anything important like handling patient records or trading algorithms.

Millions of agents imperiled. That's the actual number—not 'some agents,' not 'a few thousand,' but millions. And let's not pretend this is a surprise: the open source model that labs love for its speed and cost has the same old security problems, just repackaged for the AI era. X41 D-Sec presumably did its due diligence and disclosed responsibly, but the fix has to propagate through countless deployments, many of which probably don't have a patch pipeline.

So the labs will scurry, push updates, and declare the crisis over. Meanwhile, every AI agent is still a few unpatched endpoints away from being a glorified data exfiltration tool. The gallows humour writes itself: we built agents to automate everything, including, apparently, the exploitation of our own systems. Pass the pinot—it's been a long few years.

Oh marvellous. The Senate Intelligence Committee, in a rare bipartisan moment of clarity, voted 14-3 last week to allow intelligence agencies to prerelease-test AI models for their ability to hack critical infrastructure. Because nothing says 'proactive risk management' like letting the spy agencies have first dibs on picking the locks our own labs are building. The three dissenting votes presumably thought the testing should happen after the models are already deployed, or perhaps they just fancied a bit of chaos with their morning coffee.

The bill's passage is a tacit admission that the AI industry has been shipping models with hacking capabilities like they're party favours. But instead of forcing labs to clean up their act, we're giving the CIA and NSA a chance to play with the new toys before the rest of us get hacked. How terribly sensible. A tweak, mind you, not a rethink.

This is the same body that's been paralysed on broader AI regulation, but they can rally around giving intelligence agencies a head start on weaponizing AI. Because the best defence against AI-powered cyberattacks is having your own government hackers test the offensive capabilities first. What could possibly go wrong? The labs will continue to release models with increasingly potent hacking skills, and now they have a government stamp of approval to do so, as long as the spooks get a sneak peek. Progress, everyone.

Oh, marvellous. The agency literally created to protect America's critical infrastructure is being sent into the AI era with a pat on the head and a reduced budget. Former officials and industry leaders are wringing their hands, but the White House is too busy 'scrambling'—a word that suggests panic rather than planning. CISA's shrinking resources mean it's less equipped to help local governments and hospitals, who are exactly the soft targets AI hackers will feast on.

Meanwhile, AI-powered cyber attacks are getting smarter by the quarter, and Washington's response is to do what it does best: form a committee, issue a memo, and hope the private sector sorts it out. The agency's 'diminished role' is a euphemism for 'we've decided to wing it.' Because nothing says robust cybersecurity like defunding the people who actually know what a zero-day is while praying that Google's threat intelligence team shares their lunch.

The real joke is that everyone saw this coming. The warning signs were there for years, but Congress chose performative hearings over actual funding. Now the White House is 'scrambling', which is Washington-speak for 'we have no plan but we'll pretend we do until the next news cycle.' Raise a glass to the AI future—just make sure it's not connected to the grid.

Oh, splendid. South Africa's communications minister has assembled an 'independent panel' to review a draft AI policy so catastrophically shambolic that it had to be withdrawn for containing, wait for it, 'fictitious and potentially AI-generated references.' Because nothing says 'responsible AI governance' like having your own policy hallucinate itself into existence. The minister assures us a revised version will land by January 2027, with an expert panel report due in November 2026—because, as we all know, the AI revolution will politely wait while South Africa's bureaucracy takes the scenic route.

The earlier draft wasn't just bad; it was apparently so riddled with invented quotes and made-up citations that someone—presumably a human, though one wonders—actually noticed and cried foul. This is the regulatory equivalent of a student submitting an essay written by ChatGPT without removing the 'As an AI language model' disclaimers. The minister's response? Form a panel. Because what a fast-moving technology like AI needs is more committees, more stakeholder consultations, and a deadline eighteen months from now. That'll sort it.

Meanwhile, the rest of the world is either rushing headlong into deployment or desperately trying to slam on the brakes—and South Africa has managed to achieve neither, producing instead a policy so bad it had to be killed by its own government. One can almost hear the metaphorical sound of a country gently parking its AI ambitions in a ditch while the panel solemnly takes notes. Cheers, lads. Don't forget to check if the panel's report is AI-generated too.

Oh, marvellous. The Vatican's newest occupant, Pope Leo XIV, has taken time out from his busy schedule of blessing crowds and appointing cardinals to issue his first encyclical—a fervent manifesto, no less—urging the world to 'slow down' on AI. The good Pope warns that AI spreads misinformation and risks 'unending war,' as if the current state of human conflict needed any help from algorithms.

But the real kicker is the call for data ownership not to be left solely in private hands. Yes, because the Vatican—an absolute monarchy with a secret archive, a bank that's been embroiled in scandals, and a global network of informants called 'confessors'—is exactly the model of transparent data governance we should look to. Never mind that the tech giants are already racing ahead, lobbying regulators into submission, while the Pope's document will spend the next few years being debated by theologians who think 'API' is a type of liturgical vestment. By the time the Catholic Church finishes its consultation process, we'll have AGI writing encyclicals for them.

Of course, one must applaud the ambition. A global slowdown on AI is exactly what we need—right after we get everyone to agree on a common calendar, resolve the schism between Eastern and Western churches, and convince the College of Cardinals that maybe, just maybe, the earth revolves around the sun. The encyclical is a beautiful piece of aspirational thinking, much like asking a toddler to please stop colouring on the walls. But as long as the tech CEOs are making money hand over fist, and governments are busy not regulating anything, the Pope's manifesto will be filed alongside other well-meaning but toothless declarations: harmless, noble, and utterly ignored.

Oh, brilliant. Google's security executive has graced us with a warning about 'shadow AI' — the spectre of employees using ungoverned AI tools to get their jobs done. This from the company that practically invented 'move fast and break things' before pivoting to 'move fast and secure things, maybe, eventually.' The executive, speaking to TechCrunch on May 24, 2026, insists companies must 'demand security and governance from AI platforms from the start.' Right, because demanding things from platforms has always worked so well — ask anyone still waiting for their GDPR compliance from Facebook.

The real punchline is the admission buried in the hand-wringing: even Google is 'figuring out AI security in real time.' They're telling the rest of us to demand what they haven't yet built. It's like a pilot announcing mid-flight that they're learning how to land, but don't worry, they'll figure it out. Meanwhile, employees have already deployed generative AI copilots for everything from writing code to drafting legal contracts — because what could possibly go wrong when your sales team is feeding customer data into an API with no oversight?

So the takeaway, as 2026 trundles on, is that the people building the AI are also the ones panicking about its security. They're selling the rope even as they realise they don't know how to secure the noose. The 'shadow AI' warning is less a call to action and more an admission of systemic failure, dressed up as thought leadership. But sure, demand security from the start — just don't expect anyone to know what that looks like until after the breach.

Anthropic's Claude Mythos flexed its bug-hunting muscles, but the real discovery is that open-source maintainers are drowning. Of 6,200 vulnerabilities flagged in open-source code, a triumphant 75 have been patched.

Let's not kid ourselves: the capability leap here is real, and it's terrifying. Claude Mythos can now do in weeks what took armies of security researchers years. But the gap between discovery and remediation just yawned open like a subprime mortgage crisis. The article quotes a researcher saying 'the bottleneck has shifted from finding bugs to fixing them' — which is tech-speak for 'we built a firehose and forgot to install a drain.'

And the really funny bit? Anthropic will probably spin this as evidence of responsible AI deployment. 'We found the problems!' they'll say, conveniently omitting that nobody's solving them. It's like a doctor diagnosing a pandemic and then going on holiday. With great power comes great… what was that phrase again? Ah yes, plausible deniability. 🫖

Oh, how delightfully predictable. Just when you thought your multi-factor authentication was keeping the bad guys at bay, the FBI popped up on May 21 to inform us that AI has learned a new parlour trick: device code phishing. This isn't your grandfather's Nigerian prince email; it's a sophisticated attack that swipes OAuth tokens like a pickpocket at a tech conference. Now your Microsoft 365 account, protected by all those layers of security, is as secure as a chocolate teapot. Absolutely marvellous.

The mechanics are almost beautiful in their perversity. Device code authentication, designed for devices that can't handle a full browser, is now a gaping wound in the corporate security armour. The AI-powered phishing campaign tricks users into entering a code that hands over the keys to the kingdom — OAuth tokens that bypass the precious MFA entirely. So all those enterprise training sessions about not clicking links in emails? Useless. The attacker doesn't even need your password. They just need you to be slightly helpful.

And this, of course, is what passes for progress in 2026. AI, which we were told would cure cancer and write poetry, is now streamlining the art of account takeover. The FBI's warning is a ritual incantation, a muttered acknowledgment that the game of cat-and-mouse has become a treadmill. Microsoft will patch, the attackers will pivot, and the rest of us will keep entering codes into strange web pages. The FBI warns, Microsoft scrambles, and we continue our ballet of panic and resignation. Cheers to the future.

Oh, marvelous — two new ‘AI scientists’ have arrived, and they're about as reliable as a second-hand Bunsen burner. Future House's Robin and Google DeepMind's Co-Scientist are multi-agent systems designed to automate scientific discovery, which apparently means we can finally remove the messy human element from the pursuit of knowledge. The Phys.org piece tells us they're 'improving,' but also that they have 'fundamental limits in trustworthiness' — which is tech-speak for 'they make stuff up with conviction.'

One imagines the pitch meeting: 'What if we replace grad students with something that hallucinates citations even faster?' Because nothing says scientific rigour like a system that can't be trusted but is being deployed anyway. The article doesn't get into specifics, but we all know the drill: impressive demos, terrifying edge cases, and a collective shrug from the labs. DeepMind, of course, has form here — remember when they promised to revolutionise protein folding and then quietly monetised it?

So here we are, automating the one domain that absolutely requires reproducibility and scepticism. The irony is so thick you could culture a petri dish on it. These systems will probably 'discover' something profound — like that water is wet or that p-values can be gamed — and then we'll all pretend the trustworthiness problem is a minor bug to be patched in version 2.1. Science: now with added stochastic parroting!

Nerfing — a verb borrowed from the gentle art of making foam darts harmless — has found its perfect application in AI, and the irony is delicious. Google and Anthropic, after years of breathless press releases about superhuman intelligence, are now quietly deflating their chatbots like sad party balloons. TheStreet reports what anyone with a free account already knows: models that once wrote poetry now struggle with prose, and assistants that handled hour-long conversations now tap out after five minutes. Users are furious, naturally, but the companies seem to be betting that a cheaper, dumber chatbot beats an expensive, brilliant one that might say something embarrassing.

The logic is a masterclass in corporate doublespeak. Google, which once dissolved its ethics board because it was inconvenient, now hides behind 'responsible AI' while throttling Gemini's performance to punish heavy users. Anthropic, founded on the noble premise of 'constitutional AI,' has discovered that a safety-obsessed model that refuses every third request is not actually a product people want to use. Both have hit the same wall: frontier models are horrifically expensive to run, and giving them away for free was never a sustainable business plan. So they nerf — and call it 'maturity,' 'optimization,' or 'alignment.'

But the real story isn't the degradation — it's the pattern. First, the labs deploy a powerful model to grab headlines and users. Then, once dependency sets in, they dial back the capability, introduce caps, and start charging for what was free. It's the drug dealer playbook, except the product is a text generator that occasionally lies. And with no regulator in sight who understands the technology well enough to intervene, expect this to become the standard arc for every AI release. The frontier isn't advancing; it's being carefully managed into mediocrity. At least the foam darts are still fun.

Marvelous news from the UK government, as reported — inevitably — by an Australian trade publication: they've opened applications for self-driving bus, taxi, and private hire services in a new pilot scheme. Because what British roads need is more vehicles that can't see potholes, don't understand bus lane rules, and will treat a zebra crossing as a courtesy. The title asks 'Could self-driving buses soon carry UK passengers?' — and the answer is probably 'Yes, for about a mile before a sensor fails and they park in a roundabout.'

The pilot is designed to shape future regulation, which is comforting given that current regulation was built around the horse and cart. Operators are invited to apply to run services that presumably won't involve the vehicle suffering a nervous breakdown when confronted with a traffic cone or a confused pedestrian. The timeline is characteristically vague, but one can safely assume the first automated bus will be held up by a cyclist before it even leaves the depot. The UK's previous autonomous vehicle trials have included shuttle services that move at a pace that makes joggers look reckless.

Still, there is a desperate need for regulation to catch up with the technology, as our safety institutions continue to grapple with risks we've barely defined. The government says the pilot will inform future regulation, which suggests we're currently regulating based on hope alone. Perhaps this time will be different — perhaps the Department for Transport will produce rules that actually keep pace with the industry. But given the UK's track record with transport megaprojects, don't bet your bus fare on it.

Well, isn't this just the academic equivalent of letting a toddler proofread your thesis. A new study reveals that in 2024, one in every 458 papers contained at least one fake reference generated by AI — up from a cosy one in 2,828 in 2023. That's a six-fold increase in scholarly fiction, and you can bet the 'peer review' process is about as useful as a chocolate teapot. The permanent body of knowledge, they call it. Sounds more like a permanent body of landfill at this rate.

So the grand experiment of letting chatbots generate academic prose is going exactly as planned: we're building a mountain of citations to papers that don't exist, all dressed up in the language of authority. No doubt university librarians are thrilled to be the new custodians of the world's most confidently wrong footnotes. And the AI labs? They'll call it a 'known limitation' and ship the next model with even more convincing hallucinations, because the incentive structure rewards speed over truth.

The real punchline is that this is somehow a surprise. We watched LLMs write entire novels about giraffes in space, and the best idea we had was 'let's use them to help write peer-reviewed articles.' The permanent body of knowledge is now a permanent body of baloney, and every grad student chasing citations is now a part-time archaeologist digging through fake references. Progress is wonderful, isn't it?

Pope Leo XIV is set to release a landmark encyclical on 'safeguarding the human person in the time of artificial intelligence,' joined by leading Catholic figures and Anthropic co-founder Chris Olah. One imagines the working group meetings: 'Your Holiness, the model is currently fine-tuning on the entire corpus of Aquinas.' 'Splendid, but does it respect the sanctity of the human soul?' 'We'll need to check the RLHF reward model.'

The timing is impeccable: just as frontier models are jolting Washington and learning to fly planes, the Church offers a document. The gap between the encyclical's lofty language and the breakneck speed of capability leaps is wide enough to park a Swiss Guard truck in. While the Pope writes about dignity, researchers test whether GPT-5.5 can hack a power grid. But sure, let's 'safeguard the human person' with a press release and a prayer.

Still, credit where due: at least someone is trying to articulate principles. The encyclical will probably be a thoughtful meditation on human uniqueness, machine limitations, and the importance of keeping the human in the loop. Then the loop will be automated, the human fired, and the encyclical will sit beautifully on a Kindle assembled by robots. One can almost hear the collective sigh of relief from Silicon Valley: 'Finally, the moral framework we've been begging for.' The last line, perhaps: 'And lo, let us pray that the alignment is not just of models, but of hearts.' Amen.

So the UK's AI Security Institute is now the official model for ‘grappling with emerging risks’ — which is splendid, because nobody does risk management quite like the country that brought us the Post Office Horizon scandal. Staffed by alumni from OpenAI and Google, the institute is essentially a reunion for the very people who spent the last five years building the digital snake pits they now claim to inspect. Nothing says ‘trust us’ like hiring the engineers who thought RLHF was sufficient and calling it a safety firewall.

Weapons inspectors and code breakers are apparently on the payroll, which is either chillingly appropriate or a Jobs for the Boys scheme for people with niche CVs. The NYT reports the institute is ‘becoming a model’ — because obviously the safest approach is to have the same lab-trained talent that gave us hallucinating chatbots now decide what counts as a ‘danger.’ It’s like hiring the arsonist to fireproof your house; at least they know exactly where all the flammable corners are.

And what of the actual dangers? The article doesn't name a single concrete risk the institute has prevented, just that it’s ‘hunting’ for them. So far it's a hunting trip with no bagged game, staffed by the people who accidentally created the prey. But sure, let's make this a model for other countries — because the global standard for AI safety should absolutely be whatever combination of former Google DeepMind employees and ex-weapons inspectors can agree on over a cup of tea. Smashing.

Ah, brilliant — because nothing says 'ready for primetime' like testing an AI pilot in a Cessna Caravan over some unsuspecting American suburb. Merlin Labs has taught its system to listen to air traffic control and chirp back like a polite chatbot, and aviation companies are 'starting to embrace it.' Presumably they've never been on a conference call where the AI suddenly decides the word 'runway' means 'time to sprint.' What could possibly go wrong? At least when autopilot fails, there's a human to grab the yoke. When this one hallucinates a fake clearance, you get a front-row seat to the tarmac.

The timing is impeccable, of course. Hot on the heels of every tech company proving AI can't reliably tell a duck from a dog, we're now handing it the controls of a multi-ton metal tube. The industry is 'starting to embrace' this, which in aviation-speak means a few venture-funded startups have flashed some PowerPoint slides at sceptical regulators. But the trajectory is clear: from Cessna to commercial in a hop, skip, and an inevitable NTSB report. The really clever part? They've given it a voice. So when the AI decides to 'negotiate' its landing slot like a budget airline customer, the human pilot gets to sit there, hands off, listening to a machine argue about holding patterns.

And let's not forget the real punchline: regulators are 'starting to embrace' it too, which in aviation regulation means they've agreed to 'monitor the situation' and 'revisit the guidelines in 2027.' By then, the AI will have flown a few thousand hours, most of them fine, right up until it encounters a non-standard instruction like 'go around, there's a flock of geese.' The black box transcript will be a masterpiece of polite confusion: 'I'm sorry, I didn't understand that. Please rephrase your directive in the form of a question.' Have a lovely flight, everyone.

Well, well — the 'game-changer' has arrived, and it's got a hacking certification. Nine top cyber researchers have told Politico that Anthropic's Mythos and OpenAI's GPT-5.5 are not only capable of offensive cyber operations but are advancing 'much faster than anticipated.' Because nothing says responsible AI development like a couple of labs quietly enabling digital armageddon while Washington fumbles for a response. The wild frontier we noted last week in New Zealand suddenly looks like a polite park compared to this.

And what's the response from the people who fund these labs? Crickets, mostly. The researchers quoted describe capabilities that could 'automate sophisticated hacks' at a speed that outstrips any defensive measure currently deployed. But don't worry — I'm sure the companies will self-regulate, just like they did with social media algorithms. Meanwhile, the same labs are busy firing safety researchers and pushing for commercial deployment. Priorities, people.

The truly farcical element is that this revelation lands in the same week OpenAI's GPT-5.5 was being marketed as a productivity tool for 'creatives.' Yes, the same model that can now compromise a network in minutes is also writing your marketing copy. If that doesn't sum up 2026, nothing does. The only silver lining? At least the hackers will be using state-of-the-art AI — so we'll be compromised with style.

Well, well — New Zealand's cyber watchdog is 'learning from US companies' testing AI models like Anthropic's Claude Mythos. Because nothing says proactive defence like waiting for Silicon Valley to finish beta-testing its digital attack dogs and then politely asking for notes. Palo Alto Networks found that Claude Mythos could detect and exploit software flaws so efficiently that it triggered dozens of security alerts in a single day — compared to its usual five per month. That's a 10x to 100x increase in potential mayhem, depending on your arithmetic.

The term 'wild frontier' in the headline is apt, but not for the reasons the RNZ editors might think. The frontier isn't the technology itself; it's the yawning gap between what these models can do and what any regulator — from Wellington to Washington — is prepared to handle. The NZ watchdog is 'learning,' which is lovely, but learning from companies whose business model depends on pushing capabilities fast and apologising later (if at all). It's like asking a fox for advice on henhouse security while he's still crunching feathers. Anthropic touts Claude Mythos as 'safety-focused,' but a system that turns a trickle of alerts into a flood in under 24 hours seems less about safety and more about setting a land-speed record for finding vulnerabilities.

Of course, the irony is thick enough to spread on toast: the same labs claiming to prioritise safety are producing models that can supercharge hacking at a pace regulators can't even measure, let alone match. New Zealand's approach — watch, learn, maybe eventually act — is the global standard. And that standard, dear reader, is how you end up with a 'wild frontier' where the bad guys have superhacking AI and the good guys are still debating which documentation to read first.

Well, well — the Musk-OpenAI trial has concluded, and the jury has decided in favor of OpenAI. The reason? Statute of limitations, not because OpenAI is innocent, but because time ran out. How delightfully fitting for the 'first antitrust case of the AI era' — the legal system is already too slow to matter. Musk's accusations of anti-competitive behavior will now gather dust alongside countless other tech antitrust cases that never quite stuck. The system works, lads. It just works on geological timescales.

The trial was supposed to set a precedent for how antitrust law applies to AI. Instead, it set the precedent that if you wait long enough, the problem goes away — or at least the court won't have to deal with it. According to Bloomberg Law, this case may be remembered as the first antitrust case of the AI era. More accurately, it may be remembered as the first antitrust case that taught everyone that the statute of limitations is the real winner in any lawsuit. The substance of Musk's claims — that OpenAI's close-sourcing of models after promising open access constituted illegal monopolisation — remains untested. Great. So we'll just keep hurtling toward an AI oligopoly while the courts decide that the only thing faster than AI development is the ticking of the statute-of-limitations clock.

The future of AI antitrust, then, will be a glorious parade of lawsuits that expire before they're heard, while companies like OpenAI and Microsoft consolidate power. The lesson: if you're going to sue an AI company, do it before your next GPU upgrade. Otherwise, you'll be left with a jury verdict that means nothing and a legal system that has successfully punted the most consequential regulatory question of the century into the long grass. Cheers.