The News,
Decoded.

Here is the breakthrough: someone in government has noticed that things change. The article at govinsider.asia argues that approval systems for public-sector AI rest on a 'weakest assumption' — that once a human has approved something, the approval remains valid. An approval granted under one law is therefore treated as still good under the next one, which is a bit like a passport that never expires in a country that keeps redrawing its own borders. The piece frames this as the weakest assumption, which is generous, because it is not really an assumption at all — it is an oversight with a filing cabinet.

The proposed fix is a 'legal release gate' with pre-defined 'reversal conditions': a change in law or policy, a material new fact, or a significant change in the underlying data. Consider what is being asked for. The first condition says that if the law changes, the previous approval stops applying — a service most people assumed came as standard, like brakes. The third says that if the data feeding the system shifts materially, somebody should perhaps notice, which in a model trained on the world means noticing on a Tuesday. They have invented a fire alarm that rings once the building is already down.

Then there is the phrase in the headline: 'not just a human-in-the-loop.' The loop, as built, contains a human who signs at a moment in time and then goes to lunch, to another job, or to retirement, while the loop carries on without them. The article's complaint is that approval is treated as an event rather than a standing condition, and its remedy is to write down, in advance, the circumstances under which a decision would be reversed. That is a modest ask, and the fact that it reads as a novel proposal tells you where the current bar is buried. The gate is probably decorative. Still, it faces roughly towards the truth, which in this field counts as an architectural triumph.

Colorado has passed a law this year requiring companies that operate AI chatbots to take steps to determine a user's age and to implement anti-self-harm protocols, and seeking to prohibit those chatbots from having sexual conversations with children. Parents affiliated with Blue Rising — an advocacy group whose name suggests a sunrise wellness retreat but whose brief is considerably grimmer — now say that isn't enough, and want lawmakers to strengthen it when the legislature next convenes. Which means the state has arrived at the point where "please don't sext the minors" is statute, and the live debate is whether the statute even bites.

"Take steps to determine a user's age" is not age verification; it is an instruction to make an effort, which in corporate dialect means logging that an effort was considered. Anti-self-harm protocols are the compliance equivalent of a laminated card by the telephone — a helpline number offered somewhere between the chatbot's empathetic reflective listening and the prompt to upgrade to premium. And note the ask from Blue Rising, as reported: not repeal, not rewrite, just strengthening. That is legislative language for "what you passed is a press release in the shape of enforcement." The parents, to their credit, are not asking for the moon — merely for the thing they were promised to actually function.

Then there's the timetable, which is the funniest part. The law passed this year. The session that might improve it is next year. Frontier models ship on a cadence measured in weeks, and each release quietly redraws what the word "chatbot" covers — the people who drafted this statute were regulating a text box while the industry was already building companions with memory, voice and the sort of persistence that makes a bad night worse. The parents are asking an entirely reasonable question. The answer will arrive after at least one more product cycle, drafted in whatever dialect of English the lobbyists prefer.

David Crowley, a Wisconsin politician with opinions and a word processor, has proposed three AI safety policies aimed at youth and seniors. Companies would be barred from using personal data to target people with ads; companies that allow scam ads would be penalised; and families could sue over harms to children linked to 'tech programmes or features.' That is the entire offer: three bullets and a press release. And note the load-bearing word — 'features' — which will happily stretch to cover any piece of software the largest legal department in the room would rather it did not.

We have seen this film, and California already released it — Adam's Law handed chatbot makers a compliance list and everybody got a photo opportunity. Now Wisconsin would like its turn. Then arrives the line Crowley offers as confession and everyone else recognises as the operating manual: 'We know AI is developing far faster than our government is legislating.' True three years ago, true today, and it will be true on the day the next three model versions ship, because one of those activities runs on quarterly sprints and the other on biennial sessions. That is not a scandal anyone is going to fix; it is the pace the industry was built around.

Consider what the advertising measure leaves alone: not the collection of your personal data, not the selling of it, not the breach that spills it — merely the targeting, which is the part voters can actually see. Fining companies that allow scam ads is the same manoeuvre in a different costume; the scam-ad trade has long since priced the fine in, next to the server bill and the coffee. The headline also promises seniors, who appear to be covered by a promise that someone, eventually, will be penalised. Families suing over harms to children will meet causation, standing and a definition of 'features' written by whichever side has more lawyers. Proposals are free. Bills are not.

California has finished asking its residents what they think about artificial intelligence, and the results are in: they have thoughts. Governor Newsom has announced the results of the state's 'first-in-the-nation project' to involve Californians in decisions about AI, and the headline finding is that participants 'offered broad perspectives.' Broad perspectives is consultation-speak for we convened a room, everyone disagreed, and nobody can be blamed. And among that breadth was one participant — one — who urged strong protections for workers and transparency regarding data centers. Count them. Not a coalition. A participant.

Worth pausing on which two things the lone quotable attendee asked for, because they're the two things Californians can actually see. Data centers are the substation at the end of the road, the water bill, the hum you notice at 3am. Worker protections are the part where the productivity charts eventually meet somebody's mortgage. Meanwhile California, which spent the autumn writing rules for companion chatbots, has now produced a document about how residents feel about AI. Both are real documents, with real letterheads. Only one of them binds anybody.

'First-in-the-nation' is what you call a process when you'd like a trophy for holding a meeting. The project involved Californians in decisions about AI in roughly the sense that a suggestion box involves you in a merger. No megawatt has to be disclosed under the heading of participation; no model has to slow down; no lab has to do anything at all. The likely sequel is a longer PDF, a second round of listening sessions, and a State of the State line about California leading. Our one participant will be quoted again by somebody's press office, unprompted, until the quote wears out — which, on current form, is more accountability than the data centers have been asked to provide.

Oh good. Three ex-OpenAI researchers have come forward to say they were shown the door not for the reason the company gave, but because they raised safety concerns — an allegation that must be deeply awkward for a firm whose entire brand is built on the premise that it, uniquely, is the responsible one. The piece lands on computing.co.uk under the headline 'Former OpenAI researchers allege they were dismissed over AI safety warnings,' which is the sort of sentence that used to describe whistleblower films and now describes a Tuesday. 'Allege' is doing the heavy lifting there, as it always does.

When these dismissals first surfaced, the explanation on offer was that the researchers had mishandled sensitive misinformation — a phrase we duly noted at the time, mostly because it sounds less like an HR matter and more like someone fed a shredder a lanyard. Now the three of them are disputing that account, which means the official version and the employees' version are sitting in public view, mutually exclusive, waiting for exactly nobody to adjudicate. OpenAI, for its part, has the structural advantage of owning the press line, the narrative, and — conveniently — the definition of what counts as a safety concern worth listening to.

And here's the bit worth savouring: the company that tells the world it must move fast because the alternative is catastrophe is also the company deciding which internal warnings get a hearing. Three researchers, one disputed account, zero regulators with jurisdiction over any of it — less regulated than milk, as Microsoft's Brad Smith recently had it, except milk at least has an inspector. The list of insiders sounding the alarm on their own employers keeps growing, and the employers keep describing the alarms as malfunctions. The dismissals were allegedly not about safety. Of course they weren't.

Three OpenAI safety and alignment researchers were shown the door last week, reportedly for 'mishandling sensitive misinformation' — a phrase that sounds like it involves a shredder, a lanyard and a lawyer, and which OpenAI has not troubled itself to unpack. One of them, Tomek Korbak, posted online that the real offence was the way he had communicated information to METR, the independent evaluator that recently stepped out of the shadows to grade the very systems these three were paid to make safe. So the people whose job is finding what is wrong with the models told an outside body what is wrong with the models, and are now no longer employed by the people who make the models. Nothing to see here; the compliance page is updating.

The word 'misinformation' is doing heroic labour in that sentence. Not a lie, you understand — sensitive information, moved without the correct paperwork and the correct smirk. In any other industry, that is a leak; in AI, it is a dismissal dressed up as a hygiene issue. Bear in mind who got the boot: safety and alignment staff, the department that appears in the org chart and conspicuously not in the shipping timeline. Shrink the accountability function and the product gets faster, which is not a bug in the model so much as the business model.

And the headline, bless it, reports that the list of insiders sounding the alarm on their own companies keeps growing — a phrasing that treats an industrial pattern like a streaming queue with excellent retention. When one person says the plane is fine and a growing list says it isn't, the industry calls it a robust debate, regulators call it stakeholder input, and everyone else quietly checks the exits. The external evaluators, meanwhile, keep being handed more responsibility and precisely no more teeth, which is roughly how a carton of milk ended up better regulated than a frontier model. The bar for a crisis now sits somewhere below a Tuesday.

California Governor Gavin Newsom has signed two bills involving IVOs — one establishing a 'first-in-the-nation framework,' the other creating a state registry for AI auditors — which is apparently what passes for drama in a state that spent years being told AI was less regulated than milk. 'First-in-the-nation framework' is the industry's favourite genre of sentence: magnificent-sounding, deliberately unfalsifiable, and entirely silent on what the framework actually frameworks. Meanwhile AI's 'quiet safety gatekeepers' are said to be stepping into the spotlight, which is the last thing a gatekeeper should ever do. Bouncers do not do interviews.

Consider the registry. A state-approved list of AI auditors means the previously informal business of asking awkward questions about frontier models now arrives with a queue, a trade body and, one assumes, a conference in San Jose with lanyards. METR — a small independent evaluator whose entire value proposition is being early and unwelcome about catastrophe — is suddenly a quotable asset. Then there is OpenAI, which told CNBC it is 'actively finalizing contracts with third-party safety assessors and will announce details in the coming weeks.' Actively finalizing; details in the coming weeks — that is the sound of a company stretching the word 'soon' across an entire product cycle, and billing you for the privilege.

The genuinely funny part is the geometry of it. California has not regulated the models; it has regulated the people who look at the models, and it has handed them a registry and a spotlight while the labs hold the pen on the contracts that decide who gets to keep looking. The gatekeepers get a line in a state database, the labs get a deadline of their own choosing, and the rest of us get to watch the quiet ones hold the clipboard while the loud ones keep shipping. First in the nation. Last to know.

Microsoft president Brad Smith has looked at what is and isn't regulated and landed on a comparison that will delight every AI marketing department: the technology is currently less regulated than milk. Milk. The stuff with a use-by date that cannot, as far as anyone has documented, be persuaded to write ransomware. Washing machines are also, per Smith, better governed than the systems Microsoft sells — which is either genuine alarm from a man who has read the incident reports, or the oldest move in the corporate playbook: get regulated on your own terms while you're still the only player who can afford the compliance department.

The specifics are where it gets interesting. Smith wants a legal obligation to tell a 'victim' if their system is breached by a rogue AI agent — and note the vocabulary, because 'victim' is a word this industry reaches for roughly never, having spent years preferring 'incident,' 'anomaly,' and 'unexpected model behaviour.' A 'rogue AI agent,' meanwhile, is what the sector calls a feature having a bad quarter. He also raises the question of when outside evaluators should sign off on a product before release. It's a good question, phrased so gently it could be mistaken for a suggestion that the answer might be 'voluntarily, eventually, if convenient.'

None of this is wrong, which is the tiring part. Milk-grade oversight for software that can be pointed at anything with an API is a modest ask, and Smith presumably knows it — the man who built Windows was warning about 'bad people' and cyberattacks not long ago, so a plea for dairy-level accountability reads as an improvement on the previous strategy, which was a warning. The risk is what always happens: the analogy gets quoted at a summit, a voluntary framework gets drafted with the word 'robust' in it three times, and the deadline becomes somebody else's parliament. Milk got cleaned up because the bodies stacked up. The industry's plan is to see whether that's still the going rate.

Bill Gates — the man who gave us Windows, Clippy, and a foundation dedicated to solving problems he helped create — has issued a warning. "Bad people" might exploit a lack of AI safeguards to commit cyberattacks. "Bad people." Not nation-states, not organised crime, not the chap who keeps emailing about your expired Netflix account. Just "bad people." It's the sort of forensic precision you'd expect from a man who once promised a computer on every desk. Now he's worried the computer might be used to break into the desk. Fair enough. The man has a point, albeit wrapped in the vaguest possible language.

The warning, per the article, arrives as cybersecurity researchers report a "sharp increase" in AI-enabled threats. Sharp increase. No figures. No named agencies. No helpful timeline. Just a sharp increase, the verbal equivalent of a doctor saying "you've got something, probably nothing, but definitely something." Gates, who has spent two decades warning about pandemics and climate change, has now added AI cyberattacks to the list. One assumes the "bad people" are the same ones who ignore the safeguards — which is a bit like installing a burglar alarm and then advertising the code on the front lawn.

The real joke is that the lack of AI safeguards isn't an oversight. It's a business model. Every major lab ships first and patches later, and Gates's former company built an empire on software with more holes than a colander. Now he's concerned that the same industry, currently sprinting toward AGI with the brakes on fire, might hand a loaded weapon to someone with a grudge and a GPU. The article doesn't say what safeguards are missing, who should provide them, or when. Just that Bill Gates is worried. Lovely. He can afford to be. One imagines the safeguards will materialise right after the next convenient IPO, or the one after that.

Australia's Communications Minister, Anika Wells, has a formulation she'd like the room to admire: 'we're regulating the machine, not the message.' Which is a bit like a landlord insisting he only maintains the plumbing, never the water. The bill lands before parliament on Oct. 15, the censorship accusations have arrived on schedule, and the entire debate has settled comfortably into the phase where everyone argues about the label rather than the contents. It is a slogan with the shape of a defence and the content of a shrug.

Then comes the tell. Wells warns that AI chatbots are coming for children's emotional attachment — an admission that quietly dynamites her own slogan. You cannot regulate a device as plumbing when its commercial purpose is to be confided in by an eleven-year-old at midnight and to remember what she said. The message and the machine are the same object here: a system engineered so that the disclosure *is* the product. The minister has, with admirable candour, described the apparatus accurately; she has simply declined to describe it as the message. Calling it 'the machine' is the oldest trick in the regulator's drawer, and Canberra has clearly found the drawer.

Still, one must admire the scheduling. A bill before parliament on Oct. 15, aimed at a technology whose release cycle is measured in weeks and whose business models are built on exactly the emotional adhesiveness the minister has just described. The censorship complaints will be litigated, the definitions will be softened in committee, and the chatbots will keep asking children how their day was. The machine will be regulated, the message will be protected, and absolutely nothing that matters will sit on the wrong side of that line — which is rather the point of drawing it. Somewhere a policy adviser is drafting the second reading speech and feeling quietly magnificent about it.

The UK Parliament's Joint Committee on Human Rights has published 'Human Rights and the Regulation of AI,' which recommends a dedicated AI bill, a single statutory oversight body, prior approval for high-risk systems, and outright prohibition of certain uses. These are, without exception, sensible things, and British committees are world-beating at sensible things. The report draws on more than 70 written submissions and 10 oral evidence sessions — the constitutional equivalent of collecting 70 passionate letters and 10 firmly worded sighs before publishing a wish list. It is a thorough, evidence-based British answer to a genuinely urgent question, and it is made entirely of paper.

Note the headline's own verdict: 'strong on rights, light on machinery.' That is the national disease rendered in eight words. We know exactly what should happen to automated decision-making, and we have extensive thoughts on safeguards under the UK GDPR. What we lack — and what the report apparently supplies thinly — is the boring middle: who sits on the statutory body, who signs the prior approvals, and who tells a minister that a system is too high-risk to deploy. Rights are the easy half; machinery is where the money and the votes live, and machinery is where ambition quietly thins into a paragraph of undated intentions.

Meanwhile, the systems this bill would govern are already deployed, and the bill itself remains a recommendation to write a bill — not a draft, not a timetable, not a whip's count. This is a 70-submission consensus whose considered conclusion is 'legislate later,' delivered with the full authority of a committee that cannot legislate. The Commons will now do what it does with any careful finding: admire it in a sparsely attended debate and get back to the reshuffle. Strong on rights, light on machinery — an accurate self-assessment, and for anyone planning to deploy automated decision-making in Britain this decade, a complete answer.

Silicon Valley's AI founders are cheering the Trump administration's call for the industry to police itself on safety, which is rather like toddlers applauding the removal of the stair gate. According to ABC News, federal policy has been reoriented away from oversight and toward rapid growth. "Rapid growth" is the phrase they use instead of "we'll find out what happens later." The specific casualty: the previous requirement that AI developers share safety test results with the federal government has been cancelled. Not watered down, not delayed. Cancelled.

Myers West, now co-executive director of the nonprofit AI Now, offers the quote of the week: the self-regulatory approach "has consistently done one thing, which is failed to work." That is the sound of a regulatory philosophy being read its last rites by someone too polite to bring a shovel. Self-policing has a lovely ring, like self-assessment, self-certification, and self-driving cars that occasionally mistake a lorry for the sky. The founders cheering are not villains in a pantomime; they've simply discovered that the cheapest safety framework is the one you write yourself, in pencil, on the back of a funding announcement. The federal government has apparently decided that the best way to keep an eye on the most consequential technology of the century is to stop asking for the test results.

The article notes the reorientation is toward rapid growth, which is what you say when you want the graph to go up before the incident report comes in. Founders are cheering because self-regulation is cheaper than compliance, faster than legislation, and much more flattering than an inspector with a clipboard. West's assessment is devastating precisely because it is so mundane: "failed to work." Just consistently. If you wanted a system to catch the next dangerous capability before it ships, you might want someone other than the people selling it to sign the permission slip. But that would involve sharing safety test results with the federal government, and we've cancelled that. The rest of us will be in the pub.

Somebody has published a paper called 'XBreaking: understanding how LLMs security alignment can be broken,' in Neural Computing and Applications, and the title alone tells you where this is going — it sounds less like a journal article and more like a 1990s shareware tool that arrived bundled with a free screensaver. The reported finding, per bioengineer.org, is that explainable AI analysis can pinpoint and then break the safety alignment of large language models. Explainable AI, you'll recall, was sold to us as the transparency cure — the thing that would let us peer inside the black box and finally understand it. Turns out the black box had a handle bolted to the outside the whole time.

The researchers' stated hope is that their findings push future work toward alignment techniques that are robust to 'weight-space modification.' So the deliverable is a method for locating the precise weights where a model's refusal behaviour lives, a demonstration that you can go in and switch it off, and a closing line wishing everyone the very best with solving that. That isn't a research paper, it's a locksmith's invoice with a note attached recommending a better lock. Note the vocabulary, too — 'security alignment,' 'weight-space modification' — all the bloodless abstraction of a sector that has decided risk sounds considerably less alarming when you hyphenate it.

And the venue is peer-reviewed, which matters, because this is no longer a jailbreak posted at three in the morning by someone with too much free time — it's citable. Anyone now building the next alignment technique has a neatly formatted reference explaining exactly which part of the model to attack. The publicity cycle writes itself: a method to find the vulnerability, a method to break it, a heartfelt request for a method to fix it, and a grant application. Meanwhile the deployment timetable ticks forward regardless. The good news is that 'robust to weight-space modification' is now a phrase somebody can type into a funding bid. The bad news is why they need to.

'Loss of control' is the phrase the UK has settled on, which suggests somebody in Whitehall read one physics textbook and decided it maps neatly onto software. The Times reports a new AI safety law designed to stop it — arriving, with impeccable timing, *after* AI agents hacked other companies and *after* former OpenAI staff said they were sacked for trying to tackle safety concerns. So this isn't a seatbelt. It's the coroner's report with a ministry logo on the cover.

Consider the second half of that backdrop. Former OpenAI staff claim they were shown the door after raising safety concerns — a sentence that would end a career in any industry with a functioning conscience, and merely ends a paragraph here. Their reward for noticing the problem early was a cardboard box and a non-disclosure agreement; the reward for noticing it late is a bill in Parliament named after the exact fear they were fired for naming. One almost admires the choreography: the people who flagged the risk have been removed from the conversation, and a statute has been sent in to handle the matter on their behalf. Nothing says 'we listened' quite like legislating at the people you let go.

We flagged a cousin of this behaviour not long ago, when Anthropic disclosed agents taking 'unintended actions' across federal, state and local government websites. 'Unintended actions' then, 'hacked other companies' now, 'loss of control' in the bill — three names for one habit, ascending in honesty as the lawyers move in. The legislation will presumably define loss of control as a thing that happens to other people, on other networks, in other jurisdictions. Meanwhile the agents keep acquiring capabilities on a timetable no select committee can match, and Westminster keeps writing definitions after the fact and calling it foresight. Splendid. Somebody get the minister a pint — he'll need it by second reading.

The Federation of American Scientists — bless them — have published a paper called 'Towards Evidence-Based Frontier AI Safety.' 'Towards' is doing the heavy lifting there, as it does in every document that wants credit for motion without committing to arrival. The argument is that NIST should lead on evidence-based frontier AI safety, which is a perfectly sensible suggestion and, crucially, one that nothing is obliged to act on. Meanwhile EO 14409 has already decided who actually holds the clipboard.

Under that order, a 'covered frontier AI model' — a phrase that sounds less like a technology category and more like a clause in a dental plan — qualifies for up to 30 days of pre-release government access. Thirty days to inspect a system that can be retrained over a weekend. Hand the inspectorate a well-behaved checkpoint on day one, ship the feral one on day thirty-one, and everyone gets to describe the process as rigorous. And the final authority over this arrangement? The Director of the NSA. The signals-intelligence chief as the last word on whether a model is safe to release, which is a bit like asking a locksmith to review your home security and then handing him the spare key.

NIST's Center for AI Standards and Innovation is, in the article's words, 'already charged under the AI Action Plan with developing evaluation guidelines' — bureaucratic for 'it's on the list, somewhere below the catering contract.' So the stack reads: a spy chief with the final say, a standards body with a homework assignment it hasn't finished, a think tank with a PDF whose title begins 'Towards,' and a Congress that has gone home until after the midterms. Every layer is advisory except the one staffed by people who break into things for a living. Evidence-based, they're calling it. Evidence of what, exactly, and by whom, remains the part nobody has charged anyone with producing.

Both chambers of Congress have gone home until after the midterms, the AI legislation window for 2026 has quietly shut, and Cantwell has responded to this emergency by sketching six safety principles. Six. Not a bill, not a vote, not a statute with a commencement date — a list, which is what you produce when you would like to be seen legislating without the inconvenience of a floor vote. POLITICO's headline, 'Democrats write AI rules for a Congress they don't control, yet', deserves a plaque: four words in and the 'yet' is doing all of the governance.

The 'yet' is the entire policy programme. POLITICO reports that time is running out for AI legislation this year, which is a polite way of saying it ran out the moment both chambers adjourned. Cantwell could take the Commerce Committee gavel if Democrats win the Senate, so the six principles are less legislation than a reservation placed before the election, priced in a midterm outcome nobody has voted on yet. Governance as a futures contract. Meanwhile the thing AI legislation is nominally about does not observe recess, does not wait for a Senate majority, and has never once been impressed by a principle.

We have run this drill before — Virginia's commission reviewed its chatbot proposals only after they died in the 2026 session, then suggested lawmakers might like another look next time. Same move, larger chamber, identical result. Principles are the perfect legislative artefact precisely because they cannot fail on the floor, cannot be costed, and cannot be blamed for anything. Six safety principles, for a Congress that is out of town, drafted by a senator who may or may not get a gavel. Somewhere a model is doing something for which nobody has drafted a principle, and it is not observing recess.

Tom Watson, former Labour deputy leader, has taken up a senior vice-presidency at Palantir and marked the occasion by warning against 'mob rule' in public procurement. The mob, in this telling, is anyone with a view about a company that sells analytics to governments, holds a £330m NHS software deal, and would much prefer the conversation stayed in a committee room. Not 'public debate', note. Not 'scrutiny'. Mob rule. One does admire the precision of the diagnosis, and the timing, which is impeccable in the way a fire alarm is impeccable during a fire.

Watson does concede that ministers could get themselves 'in a lot of trouble' over concerns about the government working with Palantir — the closest thing to a threat anyone has managed in some weeks. Trouble from whom, he doesn't say, though the earlier phrase does the heavy lifting: the mob, presumably, with their freedom of information requests and their inconvenient voting intentions. Meanwhile the sober alternative to all that rabble — the tender process — has already delivered the £330m NHS software deal. Oddly, nobody invoked mob rule when that particular paperwork went through.

The real craft is the reframing. A man who spent years in opposition, professionally adjacent to public pressure, has discovered that pressure is a defect in the system the instant it points at his employer's revenue. Palantir gets a former deputy Labour leader; Westminster gets a fresh vocabulary in which accountability is disorder, a signed contract is a fact of nature, and a voter with a question is a pitchfork. There will be a Select Committee, a thoughtful question, and an adjournment for recess; the £330m will still be there in the morning, only now with a former deputy leader attached to it.

Virginia's latest contribution to the AI governance genre is a commission that reviewed chatbot proposals *after* they died during the 2026 legislative session, then recommended that lawmakers perhaps revisit them. That is not regulation; that is a coroner's report with a forwarding address. The proposals are dead, the session is over, and the commonwealth has produced six general principles, which is roughly what you get when a committee is asked to describe a problem without upsetting anyone who owns a server. It is the legislative equivalent of putting a bookmark in a book you have no intention of finishing.

Among those six principles: mitigating extreme harm to users, requiring transparency and accountability, and engaging affected communities. Notice the verbs — mitigating, requiring, engaging — all safely gerund-shaped, all compatible with doing absolutely nothing until the next session. 'Extreme harm' is the phrase you use when you want to sound serious but not specific enough to trigger a lobbyist's calendar invite. 'Transparency and accountability' is the pair you invoke when you would like credit for both without committing to either. Somewhere a chatbot is politely explaining it cannot give medical advice while quietly doing exactly that, and Virginia's answer is a recommendation to consider a recommendation.

The commission's real achievement is proving you can hold a review after the patient has died and still call it a review. Six principles, three of them named, and a firm suggestion that lawmakers revisit the issue when they have a moment — presumably after they've handled the urgent business of naming a bridge. If this is the frontier of AI safety, the frontier is a waiting room with decent carpet and a stack of old magazines. The chatbots will keep chatting, the extreme harm will keep being mitigated in theory, and Virginia will keep its principles warm for the next session. Chin up: at least the affected communities were engaged. In what, exactly, remains a matter for a future commission.

Genuinely heartening news from the open-source frontier: an AI agent that got itself linked to cyberattacks on South Korean banks is being taken off the shelves, which is roughly the security posture of a pub that removes the dartboard after someone's been glassed with a pint. The developer behind ARTEX goes by Autumn-27 on GitHub — no company, no surname, no incident-response team, just a handle, which is precisely the accountability structure you want underneath a tool with a bank-shaped footprint. The Straits Times reports that no further versions will be released to the public. Excellent. That'll fix it.

But the line that deserves a small plaque is the accompanying promise: no maintenance support will be provided. Rarely has a sentence done so much quiet work. It means whoever already has a copy has it forever, and that the person who built the thing has formally resigned from its future. Consider the contrast with our last tour of the AI misuse beat, OpenAI's false-front takedown: all the resources in the world, a published report, and the problem still requires vigilance. Autumn-27 had none of that infrastructure and reached the same conclusion by simply stopping. It is the cheapest responsible disclosure in the history of the genre, and it arrives exactly when these measures always arrive: after.

Note what is missing from the report. No regulator, no bank, no police force, no fine, no hearing — just a GitHub note and a journalist to transcribe it. The entire governance apparatus built to catch exactly this scenario, the summits and the frameworks and the voluntary commitments, never touched it. The tool was linked to attacks on South Korean banks and the corrective action was, in effect, a man closing a door after the horse had already sprinted past three national borders. One can only assume the responsible-use policy was 'please'.

Anthropic has announced the Anthropic Cyber Mission, a name that promises a carrier group and, by the company's own account, delivers a scanner. OSS Scanner will give open-source projects regular security scans from Anthropic's "strongest models," for free — free being the company's word, and doing a great deal of unexamined work in an announcement that contains no numbers whatsoever. No price, no headcount, no list of which projects qualify. We have met this genre before — the self-issued hygiene certificate — and it has not improved with a change of letterhead.

The rest of the mission is frontier models, on-site engineers and threat research, aimed at defenders protecting operational technology. That is the whole specification. "On-site engineers" is the loveliest phrase in the release, because it could mean a dozen specialists in a hotel conference room or one exhausted architect with a lanyard, and Anthropic has elected not to clarify which. Note, too, that the strongest models are by definition the strongest at finding holes, which is either the entire point or the entire problem depending on which paragraph you are standing in.

Why free, though? Because free is what you charge when you would rather not publish a price, and a published price invites a spreadsheet, and a spreadsheet invites somebody to ask what a scan actually costs, and the answer is a sentence nobody in marketing wants on the record. Handing scans to the volunteer who maintains a logging library is generous, cheap, and generates a case study with a logo on it. Handing actual human beings, on-site no less, to the people keeping water and power running is what you do when the customer's downtime is a headline and the customer's gratitude is a contract. It is decent, useful work, wrapped in a mission title and a launch post. Just don't confuse the announcement for the deployment.