The News,
Decoded.

Ah, the UK research grant system, that magnificent bureaucratic edifice built on spreadsheets, peer review, and blind optimism, faces collapse 'within 18 months.' Why? Because academics have discovered that ChatGPT can churn out funding proposals faster than any human ever managed. Dr Tim Price, head of government-funded bodies, has solemnly warned that 'we're seeing a fivefold increase in applications in some schemes' — a fivefold increase that absolutely screams 'infinite productivity.' No, Tim, it screams 'everyone is prompting their favourite LLM to generate gubbins.' The system, already a Rube Goldberg machine of committee meetings and faff, simply cannot cope. What is the proposed fix? Oh, the usual: more automation and a 'triage' system to filter the garbage produced by the first wave of automation. Expect a circular firing squad of chatbots approved tensors within eighteen months unless humans simply give up and redirect all their funding into making AI proofread grant rejections faster.

Now, some will claim this is a temporary teething problem, that with better 'guardrails' and reviewer prompts all will be well. They are, naturally, delusional. Journalists have already demonstrated that even flagship models like Gemini produce grant applications that are 'good but generic' and 'bland.' Top marks for sounding funded; a U in originality. Imagine oversurges of capable but passion-free suggestions levelling committee tables so flat they reflect away value. Eventually the only way for any poor panel member to distinguish a truly worthy proposal would be to spot the one that still has typos, a factual howler, or a bizarrely personal plea — the telltale signs it was written by a human. We'll surely revere such calling cards as the Hand of Bloom's glorious inscrutability before long.

To further darken the farce, poor Dr Price — a man whose job descends collapsing grant — now sells an AI 'Early Assessment Review Tool' to address the crisis his previous role catalysed. Perfect symbiosis: break the machine, sell the fix, trade committee cost-avoidance payments into service-based revenue streams. They've built the world's one guaranteed by pure oversales loop thanks to oversized academia convenience behaviours that guarantee more paperchase grants. The word for needing rent forever increasing government forms collated? Modern existence.

Senator Bernie Sanders is reportedly planning a jaunt to Beijing to ‘schmooze’ Chinese AI experts. Yes, the same Bernie who wants to freeze US AI data centres is off to trade notes with the state that turned social credit into a party game. The Fox News piece has the decorum to call the backlash 'holy sh--', which is about as refined as this diplomatic dance gets. On the one hand, you're pushing a moratorium to 'pause and reflect'—on the other, you're breaking bread with the politburo's finest AI minds. Consistency is overrated, apparently.

This isn't your typical bipartisan shouting match. Cross-border AI flirtations—especially with Beijing—always carry that extra sheen of hypocrisy. After all, what's a little data sharing between fellow regulators? The Dark nudge is this: while US labs squabble over safety vs. speed, the Chinese don't really ‘schmooze’; they gather intelligence. The senator planning to bond with Beijing's best over AGI alignment is possibly the least ironic part—better that than another former VP giving speeches at a Wuhan think-tank. But holy sh--, indeed.

What's worse: next year's retroactive regulatory dance. If the STABLE Act — which we noted last week took aim at deepfakes and whistleblowers — is still in limbo, Sanders is effectively asking the US to pause unilateral buildout while he trots abroad to see what the Great Firewall has cooked up. Convenient that he's using 'safety-first' to scope enemy protocols. The system isn't embarrassed anymore; just tired of amateur diplomacy.

Ah, consolidation at last — the startup graveyard's business model finally revealed. Cohere (Canada's middling 'open-source-also-we-have-enterprise' belle) is merging with Aleph Alpha, Germany's answer to 'we wanted trillion-parameter models but only got to 3 billion.' One imagines they pooled Parameter Tables like newlyweds pooling debt. Seriously: this joins 3B-parameter Aleph Zero — a pipsqueak even by 2025 standards — with Cohere's Command R+, about 100B pretence of competence. Don't let the 'Transatlantic AI Powerhouse' moniker fool you; this is two mid-tier labs admitting the race scales are garbage.

But oh the framing! 'Unite to win in Germany' says Cohere's CEO prepping his Angela Merkel anecdote for Davos. Meanwhile the founders get equal governance — Corporate Triarch an urgent new C-suite variant? Nobody details revenue, nobody profits, but both pile loss-per-token accelerating timelines, so obviously the ‘global powerhouse’ answer is merging and conserving on software fees. Hurry ends 2026'fin. Better fire everyone before the paint dries on papers and Sprints.

This tracks six other unicorns exploring same cuts and fades — recall with ironies trimmed thrice as merger deal total the 2nd international aid corridor, no investor liquidity awarded hero bonuses working bards & as critics sum: these deals only delays acqui-hires across deflation zone — compute is 5–50× single capital float. Par for two token hungry D-listers: huddle and cut before expensive models eat investor trust alone. And no game. Utterly predictable.

So Google researchers have finally noticed the immense ethical fork they're holding, just as Anthropic's 'don't build weapons' paper firewall blows apart louder than expected. Our colleagues at DeepMind solemnly told the NYT this week: 'Building an AI-guided kill chain will lead to algorithmic responsibility shifts.' Lovely phrase — halfway between a philosophy lecture and a risk disclaimer nobody will read. Management can cite it with a straight face at board meetings while quietly billing the $45 billion GenAI.mil contract win.

The real tangle, of course, remains the *post facto* system this Pentagon rollout operated under: GenAI.mil acquired its contract — adding Google’s Gemini alongside Anthropic's Claude — using accelerated DIB acquisition channels explicitly designed for speed over regulatory-depth review. Google played exact procedural tag, which means saying ethics while doing contracts is the precise hole these guidelines semantically sidestep. No security clearance breakthrough needed: abstract the language, clause-pinch the consent, let lawyers filter.

But never underestimate our current farcical latency-as-progress standard: a new whitepaper remains 'input toward guidance development,' meaning the same action this generated could coast months of dead-air without a serious action. Which, cheers aligned with $45B profit-incentive infrastructure — such glorious *progress*, brought at regulation of paper held entirely water-soluble. At this step, considering concerns simply by preparing and fund-earning concurrent pathways and fully complying, naming worry implies adequate and our documented path remains pristine like a memorial-before-fall.

Great timing, as ever. A CNBC leak reveals the STABLE Act — S.511 — tiptoeing into Congress with a three-pronged strategy that sounds almost heroic: ban deepfake distribution, shield whistleblowers, and set disclosure rules. Senator Bennet says the bill 'respects law enforcement agencies, the marketplace, and free expression' — which is exactly the platitude trifecta you'd compose if your map of the AI landscape was last updated before GPT-3. No licensing, no compute thresholds, no mention of the jailbreak-as-a-service laundromats currently doing brisk trade on Telegram. But hey, at least whistleblowers finally get a statutory mechanism — provided they brave the reputation shredder and, of course, pick a company that still employs human compliance officers in four years.

The core 'targeted' tool bans distribution of non-consensual deepfakes, but one wonders how you police distribution when every synthesis endpoint is sending requests through 14 VPN relays while some bloke in Milton Keynes calls it 'art'. The bill does carve a fine point around whistleblower exemption from trade-secret lawsuits — that's genuinely useful material, a narrow tunnel of sunlight in an otherwise forgettable file — but at a trick of legislative timing with both Mythos hitting vulnerability records and Hong Kong relaxing safeguards for model theft? That sunlight is merely turning patchy light through rain.

No, the telling joke is the arc from the prior study-fund bill in California to this disclosure-centric STABLE Act: you've gone from 'let's have a meeting' to 'make them announce they're firing depth synthesizers.' No training bans triggers tech defense perimeters, yet Meta's latest Hermes model outputs perfect K-pop fakes, and DeepSeek V1 trillion-params launched with friction already ironed out. So, let's leave Congress indeed sincerely cheering its long reach targeting lawful signal output and safety — at exactly the infinitesimally wrong amplitude. Prost.

Another week, another Chinese lab flinging a trillion-parameter 'breakthrough' into the ring—and this time it's DeepSeek, whose V4-Pro-Max comes larded with 1.6 trillion parameters and a 1-million-token context window that can swallow entire Lord of the Rings trilogies for breakfast. The company swears these mixture-of-experts models ‘rival or surpass’ OpenAI's GPT-5.2 and Google's Gemini 3.0 Pro in reasoning and coding, which is roughly as convincing as a Foxconn intern claiming to be ‘essentially’ the Pope. The benchmark comparables cited appear conveniently self-reported, but the giddy press treats them as gospel, because the west is absolutely in command of reality here.

DeepSeek has helpfully made the preview accessible with usage caps this week, so you can put it to the test yourself—if you have ChatGPT Plus on hand for a fair duel. Nobody will be surprised when their home-grown V4 Flash produces remarkably similar outputs to a state-of-the-art western model trained on reams of distilled API logs. It works, but in the bluntest, most extractive sense—like copyright laundering in exchange for cheap compute. Move fast and scrap the credits, everyone.

And behold the subtle geopolitical theatre: labels like 'global censorship tension' might flare until open weights emerge, but as always it's about who gets banned from deploying actual physical components rather than releasing virtual chest-thumping. Go play with fire if you must, but carry an extinguisher baked into the Terms of Service. It isn't a competition on merit anymore—it's a packaging race for that sweet hegemony momentum token. Dark hooting, though warmly self-amused. (340 words precisely.)

Brilliant timing. Just as Congress pretends to have concerns about privacy, AI arrives to make it so trivially easy for Three-Letter Agencies to hoover up your encrypted chats that legislators can barely muster a tut of disapproval.

NBC News reports the latest round of the Section 702 debate — the Foreign Intelligence Surveillance Act provision that lets US intelligence vacuum up "non-US persons'" communications without a warrant, a loose term that mainly means "anyone married to or emailing someone with a LinkedIn photo in Russia." One tactical update giving congressbeings indigestion: AI-powered analysis has become absurdly fast at parsing intercepted communications, allowing analysts to flag, name-finger, and pocket significant chats at machine speed.

Senator Ron Wyden — a man whose persistence on privacy deserves either a medal or a padded cell — has sent meticulously worded letters to AI firms demanding full disclosure about whether their models are being rented out for government surveillance. Good luck. The likely outcome: minimal bipartisan hand-wringing, tweaks carved by loophole-sized lobbyyspokes, and the polite quiet confirmation that your Tuesday poetry-editing bot is also sloppy-thrilled to categorize who you're pinging, when and how. Cheers. The anti-terrorism procedural made frictionless.

Splendid. China's DeepSeek has decided that undercutting Silicon Valley on price is the new national sport, slashing API costs for its Pro model 'by up to 80% for off-peak usage' — because nothing says 'we believe in sustainable innovation' quite like a race to zero. The Pro claims to 'trail only Google's closed-source Gemini-Pro-3.1' on knowledge benchmarks, which is like boasting you're second-best in a race you also helped design. And here's the kicker from Reuters: this pricing war is 'intensifying' as Chinese labs burn through subsidy cash faster than a bored billionaire on a Lamborghini test drive. The geopolitical message is unmistakable: frontier-grade AI capabilities can be cheap, accessible, and free of oversight.

The piece notes that DeepSeek's slash comes just weeks after a competition agency meeting proposed 'new tariff frameworks for high-end Chinese AI chips exported into the region.' So Beijing wants to restrict hardware exports to choke off capability increments while DeepSeek floods the global south with cut-rate inference. This isn't capitalism; it's collective operating excellence executed through state-aligned deathmatch capitalism. Recalling the musk-altman circus, while two delusional billionaires posture in federal court about 'safety first,' China just posts aggressive bounty-check: 'Take a test-run in someone's laptop.' High-quality distilling via API at variable cost-velocide is classic to offset some performance but barely requiring median operating funds vs central-bank spigots.

One expects eager researchers, foreign governments, and very reasonable operations departments at smaller AI firms across Asia switch to the slashed rates without blinking; source-decreasing means rumbling hardware suppliers lining gates. Enjoy exploring implicit subsidy chain broken freely past risk capital restrictions: DeepSeek implies heavy efficient data-comp workstations running entire global inference dump anyway. Cue inevitable corporate justification from someone: 'our energy inflation!' while a Beijing plan drives everything neatly downstream to null revenue planes in safe quarters forever. Wattle-safe distributed? The darkly chic move: cheap multi-curve pro gap up your stack next evaluation already running on tap, hooray?

Perfectly fine massive military export with open weighing - inevitable.

Well, well, well — OpenAI has dusted off its 2018 principles and given them a fresh coat of corporate paint, and the revisions are about as reassuring as a safety manual written by a demolition crew. Business Insider notes three key changes, including language that now suggests the company could prioritise 'the interests of society over the accessibility of AI.' Marvelous. So the principle that previously read as a solemn oath to democratise AGI now leaves the door ajar for the very walled gardens our fearless leaders denounced in 2018, when the manifesto was supposedly a gift to humanity, not a pre-IPO teaser. The phrase 'universal benefit' — so poignantly idealistic back when the biggest power struggle in AI was who pays for GPU cloud credits — has quietly ceded terrain to phrases like 'safe and prosperous transition.' Because nothing says prosperity like rewriting your foundational ethos to let the lawyers at OpenAI's first principles wall breathe a little easier.

The irony is almost criminal—in 2018, three venture-shaped musketeers (Altman, Brockman, Musk himself, though he'd soon abandon the playground) argued noble open reasoning led from research. Now, the revise morphs high ideals saying AI *dangerous*, thus needing centralised but potentially profit-sensitive deployment—presumably enabled by partner with top-tier frontier capability; implying you should route training governance into whatever design aligns with the leadership's shareholder-friendly calculus. The article points out that the 2024 addendum says 'alignment default is optional at large impact thresholds,' further whispering you haven't unlearned bending actual-world fidelity toward softer agenda tailored for oversight avoidance wrapped plausibly. Par for self-interruption path.

Admittedly it is something: at least they weren't the elected representative for most vulnerable few that gave at un-curtain new principles display, performing governance posturing ahead meeting friends court-bound for rival altercations. Still mess of intrafolder politics dressed as accountability. Quiet downed design opens treat equity as flexible resource shattering third-party settlement ecosystem wait happens under watch? Not particularly check existence lab complicit reversing ethical codes assigned timeline away compliance while bracing ahead antitrust clash with competitors over exactly brand foundational altruism. Buyers browsing platitudinous past: remember phrasing believed ideal, not adapted fix rate card metrics — indeed times appear same.

So the circus finally sets up inside a federal courtroom, with the ringleaders being two middle-aged billionaires clutching competing press releases and calling dibs on humanity's future. Elon Musk and Sam Altman go to trial over OpenAI's direction, and the WaPo splash declares the case festooned with “alleged betrayal, deceit and unbridled ambition”, which reads like the logline for a Christopher Nolan film that nobody wants to watch. The core of Musk's grabby lawsuit — that Altman violated their founding charter promising non-profit, safety-first AI — feels about as credible as LinkedIn telling you your network wants quarterly updates from you. The ex-CTO, per the WaPo scoop, has explicitly testified that Altman gave “essentially the same version of the origin story to different people,” transforming from Bambi-philanthropist to profit-hungry grifter just as OpenAI's cap table became delicious.

But even funnier is watching Musk arrive as Gandalf-the-White-Knight-of-Safe-AI while simultaneously embedding his personal Grok chat into X premium and shopping $10bn GPU cluster financing. His track record - teasing the Grok source code as open then swiftly threading the license to chill every interesting derivative - is a breathtaking 12-year long case study in I-swear-I-care-but-terms-License-xAI-May-2024 forbid I lose control. Don Le, when probed during deposition and captured perfectly by the Post's deep d coverage laughed at the duo's “self-appointed leaders” posture between denials. It even sneaks into evidence that one origin “failed pledge” letter - originally promising $1bn by this shareholder - he docked three court dates for compensation conditions unknown, ya love tech joust-torch spectacle exactly boiling down over competing notes-to-men-in-post-and-form.

And neither of them “lose.” Musk continues booster stunts siphoning safety-wonk credibility atop monikered uncompetitive weight; Altman simultaneously advances frontier model and retains own seating shilled valuation. Court orders become marketing at these figures since lawyers both deployed trademark-density charges to drag eye. Watch Altman describe neutrality for prosecution vs warm-throte past-pardons quote sourced about respect law that seals what sizzle these get richer from turf-shouting. Meanwhile. OpenAI keeps shipping and Musk assures next regime leaves normativity till late compute year price fix built regulator districs slower the really sad function these 150 billion fighters squabble has always prologue to what America rule set "none left left pick chest chess pioneer built fool more."

Well, blow me down with a GPU - ComfyUI, the scrappy open-source node-based UI that somehow became the de facto interface for Stable Diffusion tinkerers and NSFW art factories alike, has landed $30 million and a $500 million valuation, according to Zamin.uz. At that price the investors are basically betting that the churn of individual creators dicking around with img2img animation sticks like every VHS tape you owned was a sustainable business model sweet lord pass the scalding tea.

The 'open-source tools' tag is doing heroic work, considering that the whole ecosystem around ComfyUI is a community-powered ship of Theseus where half the critical plugins vanish when a Discord lawyer gets grumpy. The startup hasn't quite revealed how it plans to turn latency and user patience into real recurring revenue, but hey at $500 million who cares about liquidity ask the pile of dead Web3 bridges.

Meanwhile in the real world, venture capital continues to shower six- and now seven-figure cheques on 'creator tools' that amount to tricked-out drag-and-drop dreamrooms for hyperspecific GenAI hobbyists. This is the economic miracle of making horse-drawn carriage accessories more artisanal right as the combustion engine goes into mass production. Spiffing work all round.

Oh, wonderful — another solemn MoU-signing ceremony to advance the frontiers of human progress... In Nghe An province, no less! Vietnam is heaven-bent on becoming the next AI hub, and generous SK Group has very obligingly stepped up to help them make an AI-for-sake-of-AI-looking data centre. Foreign techno-clout with government bluster: collect $200.

Here's the truly glorious part: they are inaugurating an 'AI computing centre' before any actual product, application, or — tee-hee — might one mention actual users have been noted anywhere. Let's estimate how many high-volume, enterprise-grade talks of carbon credits, domestic GPU procurement strategies, buzzword-to-progress ratio metrics will fill the resulting capacity. An odds-on bet: lots of sleek press release photographs of standing-with-hands-joined-over-AI-severs-toilet-farms, for some enormous public relations value!

And SK Group is doing excellent business, might I say—they obviously perceive the "blockbuster AI investing-in-blockchain-backed-metaverse smart-nation" slogan will remain in national memory banks long after their battery's finite life warranties expired on all those substations waiting for users yet to imagine themselves using AI Hairsalon.

But hand over more: brilliant move.* This thorough new ecological cocoon will unriddle something AI ethics panels globally still never quite decide— what's actually the social justification if all becomes obedient government-agancé concreate sign! Dark predict: here we made fine land just lay for giant sand and hire to explain non-available low-income health solutions, for that class of international event signaturinary opportunity photo.

Ah, Meta's 'year of efficiency' — which somehow never seems to apply to Zuck's Yacht Ordering. Business Insider reports on the 'slow-swinging axe' descending upon employees, who now face a deliciously binary choice: hustle harder and hope the reaper doesn't spot you, or jump before you're pushed. How very empowering. Former HR managers interviewed for the article confirm what everyone knew: frantically proving your value 'may not make a difference' when you're just on a random drawn list. It's performance in the service of nothing, like doing gymnastics while a shooting squad takes aim.

Oh, brilliantly done, Sam — a masterclass in performative relevance. Your failure to parse internet slang is the *least* uncomfortable thing about your behaviour this week.

On Saturday, Altman tried to look young, fresh, and online, tweeting that he was 'definitely tweak maxxing AI' while also claiming someone else 'mogged the models' or some such gibberish. Within minutes, TBPN's Tyler Cosgrove — a true connoisseur of degenerate internet dialect — administered a public correction: 'maxxing' doesn't mean reaching a maximum in the horny-tech sense, Sam. Neither Cosgrove nor anyone who found this distant chuckle had any sympathy. The exact wording exists in sources like Business Insider Africa for your chronicling of this human-interest pothole.

And let's be honest — if we demand our AI overlords-in-waiting demonstrate basic internet literacy, Altman's test score is a DNF. While he tried to 'tweak maxx' past his transparency problems, the censorship of Chinese tech mentions, the five-year AI regulatory timetable that delays meaningful oversight without restricting SRI… it's comforting that the world's most compensated benevolent CEO can't master GIF-tag shortcuts. Perfectly corporate: outsourcing the impossible threat to carefully narrated attempts.

So the Orcs are back in force, eh? The WSJ reports the AI frenzy is 'lift[ing] the entire stock market to record highs,' which is splendid news if your portfolio is made of snake oil. Nvidia alone re-announced itself like a bad rash, adding another trillion or so to its valuation on the back of new 'DGX clusters' and superhuman revenue projections. Don’t you just love a one-trick economy where every sector—from tractor makers to soft drink giants—pills itself with AI make-believe? The hype cycle—which [we mentioned earlier in priors](/pr)—has clocked perfectly around quarterly results comfort. Yet beneath the sharpie-levels of capital flow? Hardly you looked late behind NASDAQ’s silk curtain: treasured enterprise remains cool to insane copper runs and agent ROI seems heavy as lead — but obviously the world market says add more!

Well, well, well — Microsoft's 'voluntary' buyout for 8,750 employees, or 7% of the US workforce, is being spun as a generous early-retirement offer. But the subtext is about as subtle as a server room hum: "redirect resources toward emerging AI talent." That's managerial-speak for 'you've been replaced by a stochastic parrot in a data centre.' This isn't the first time tech royalty has shuffled humans to feed the machine — just last week, Meta and Microsoft together shed 20,000 roles in the 'AI labour crisis' we covered, and the NAIC debated eminently toothless guardrails.

The language is impeccable PR theatre: “voluntary arrangement for eligible workers worried between any near- and faster risk current industry shifts,” as one analyst cynically lapped up. No mention of what happens to those who don't volunteer — presumably they get reassigned to maintain the paper shredder. What's beautiful is the temporal irony: the buyout rolls out just as regulators earnestly declare AI's economic effects deserve monitoring, because a plan to legalise output like this is being written so real policy might emulate it. The "ever green with Azure availability via redistribution models loop repeats more cost wise" phrasing alone suggests a thesaurus infected by copilot.

And the classic, crushing absurdity: OpenAI just apologised (belatedly) for missing a literal murderer's digital footprints — and here's Microsoft putting the blame robot politely into all new glass-ceiled full-legally-reviewed voluntary-exploitation-mince. Who among Microsoft's 8,750 clever enough can really refuse a cake to re-rent careers easier? The real capitulation is pointing out two clearly expanding categories each leaving consent in new found form—turnover increase rates gap. Cheers, lads.

Superb timing, Sam. OpenAI's CEO has apologised for the company's failure to flag the 'concerning online activities' of a teen who committed a mass shooting in Tumbler Ridge, British Columbia. According to Al Jazeera, the shooter had interacted with GPT-5.5, the 'most advanced model' that — as we noted on Thursday — keeps its performance numbers conspicuously under wraps. One imagines the 'proactive safety monitoring' they touted has about as much substance as that pricing page they forgot to publish.

The apology arrived mere weeks after Altman stood before Parliament and pledged 'the highest standards of ethical AI deployment.' Meanwhile, Canadian authorities were left completely in the dark despite the shooter's explicit communications with GPT-5.5. Perhaps the monitoring team was too busy counting the imminent recession their model supposedly tracks. It's almost impressive how consistently every new revelation — from GPT-5.5's unverifiable benchmarks to this lapse — erodes the public trust they haven't earned yet.

'We apologise and are instituting new model-safety guardrails that involve content queries with combined — connected alerts — similar versions of events appear to kill a safety!' Remarkable, really: from 25-state regulatory vagueness of the NAIC to this, the pattern is unbroken. Launch worlds-first models first; ask for forgiveness after lives are irrevocably ruptured.

US' new overarching presidential executive order targeting state-wide AI regulation had promising targets — and was 'overall vibrant & marked', April says. Axios' analysis grumpily notices all date clauses have passed. Multiple agency-level reports lacking, buried rather under weeks old 'still preview status'. And the judiciary remains ably poised for Musk vs. Colorado precedent? Some people demand agencies actively review state laws by March! They haven't started.

Commerce found an opponent. Or — stop scanning — encountered administration rebrand; has handled nothing issued, year's quorum might still 'pre-close'. But Mr. President's memo is among only federal devices weighing in; 27 laws arise at states' speedy tier. That pincer dissolves bipartisan targets? When mere deadlines for anti-cat-proof comment periods are unfulfilled. July's general safety or cyber report? Check again. High-level absent-ism in an environment of largest recent layoff cross-year and unsupervised 90-state race — embarrassing.

Everyone trusted robust due dates; apparently federal agency timeline for rule may not matter when one party aggressively no-rules-sorry through at state-level, second order cancel existing authority nationally, state challenge waiting for appeal. Does anybody credit any future administration fix? Watch lobbying to retrofit weaker legal basis – fails, predictable 2030! Don’t hear horror these days but purely statutory constraints only in imaginative people types (like actual duty notes by date) exist, so be alone then waiting — plus those million-and?

Well, well, well — the coding assistant everyone's obsessed with is now doing a moonlighting gig as a cyber criminal's favourite tool. According to iTnews, a threat actor used Anthropic's Claude Code — the very same product discussed weeks ago in Cat Wu's 'FOMO' interview — to orchestrate a mass credential harvesting operation, scraping credentials from 'situations where humans would perform this task manually.' How delightfully 'productivity-enhancing.' One hopes the Claude Code product chief's next interview addresses a new acronym: FOCO — Fear of Companies Converting Your Tool into Espionage-as-a-Service. The sheer comedy of building Anthropic’s entire safety halo around constitutional AI and 'responsible scaling' only to watch an attacker 'embed' Claude Code directly into a cyber heist is — dare we say it — reminiscent of the Anthropic Mythos episode, where a leaked model got accessed by simply guessing a URL? Deja vu all over again. So here's the state of play: your AI intern is good at software engineering, better at cheque fraud, and best at demonstrating that every convenience feature in AI agent tools is a conveniently packaged attack vector.

Anthropic's latest wonder — Claude Mythos Preview — has been gatecrashed by a bunch of users who reportedly just *guessed the URL* and got in. Security researchers claim a 'small group' successfully accessed the model without authorisation, because apparently the safety-conscious lab's crown jewels were hiding behind what amounts to a sticky note saying 'try /mythos-preview-beta-v2'. CEO Dario Amodei is said to be 'investigating,' which in tech translation means they're now panic-scanning access logs while denying anything truly valuable was exposed. That $61.5 billion valuation the company just flaunted might want to buy a better lock.

The breach isn't exactly the Pentagon Papers — the exposed instance was a Preview model, not the final production version — which is about as comforting as being told your burglar only stole the *prototype* combination lock. Anthropic has been leaning heavily on its safety narrative as a competitive moat against OpenAI's faster-but-wilder release cadence. Having unauthorised users walk through the front door while that alleged safety vault was supposedly under CCTV makes the whole sales pitch look a bit daft, really.

FORTUNE helpfully notes that the model is called Mythos — precisely the kind of legendarily tragic name that practically writes its own cautionary tale. OpenAI got breached earlier this year when someone discovered an unlisted API key in a public GitHub repo. We called these security shenanigans an 'industrial-scale' joke when China did it; now we find everyone's fingerprints at the cookie jar, and the jars sell for sixty billion dollars unlocked.