The News,
Decoded.

OpenAI has announced that its GPT-5.4-Cyber model and Threat Analysis & Containment (TAC) platform will shield financial giants from AI-driven threats, and honestly, who wouldn't feel safer with the same company that spent years testing its products on paying customers now offering security services? The rollout comes conveniently accompanied by warnings from Palo Alto and CrowdStrike CEOs that unprotected firms face imminent AI exploits, as if the cybersecurity industry needed another existential threat to hawk software against. The whole arrangement feels like a fire sale from the arsonist's guild — OpenAI builds the unstoppable AI, then sells you the digital fire blanket.

The specific partnership details are about as vague as a modern art gallery's price list. OpenAI is deploying GPT-5.4-Cyber and TAC for 'several large banks,' but the press release stops just short of naming any actual victims. TAC supposedly provides real-time threat analysis, which presumably uses AI to detect the exact types of exploits that OpenAI's other models could theoretically generate — convenient! It's the corporate equivalent of selling both the lock and the skeleton key, with a straight face.

The subtext here is deliciously grim: the very firms that develop frontier models are positioning themselves as the only people competent enough to defend against the threats those models create. Given the industry's track record with safety guardrails, one might want to check that the 'secure' platform doesn't have a quiet update path for future vulnerability generation. Still, the fees will be lovely, and the next crisis will certainly generate even more consulting revenue. It's almost as if the tech industry has discovered an infinite money glitch — break windows, sell windows, repeat ad infinitum.

So ASIC has joined the international chorus of regulators 'monitoring' Anthropic's Mythos model for potential banking system risks, which is the regulatory equivalent of watching a car crash in slow motion while politely discussing the weather. The Australian Securities and Investments Commission confirmed it's keeping tabs on a model specifically designed to 'identify and exploit zero-day vulnerabilities' — a delightful feature set for financial stability — alongside its peers worldwide. One imagines the regulator meetings involve much nodding and note-taking as they track a capability that could theoretically dismantle the very systems they're meant to oversee.

This comes mere weeks after we noted the NSA had reportedly joined the exclusive club of undisclosed Mythos recipients in our previous coverage, using it to scan environments for exploitable flaws. Now we have financial regulators monitoring the same tool, creating the charming scenario where intelligence agencies are actively weaponising vulnerabilities while their civilian counterparts watch from the sidelines with concerned expressions. Anthropic's 'lay the risks out on the table' approach appears to have resulted in regulators gathering around said table to admire the fine craftsmanship rather than actually removing the loaded gun placed upon it.

The sheer bureaucratic inertia on display is almost impressive: a model with explicit vulnerability exploitation capabilities is being deployed to intelligence agencies while financial watchdogs 'monitor' its development. No bans, no restrictions, just careful observation as the tool that could theoretically collapse banking systems continues its rollout. One wonders what exactly constitutes sufficient evidence for action — perhaps when the first zero-day exploit actually triggers a bank run, rather than merely having the capability to do so. In the meantime, regulators worldwide will continue their diligent monitoring, presumably updating their risk matrices while the actual risk matrix expands exponentially.

So Anthropic has decided to 'lay the risks out on the table' with its Mythos model, a delightful bit of corporate theatre that suggests they've discovered a new piece of furniture for displaying their wares. The chief of relations with startups assures us this transparency is 'central to Anthropic's DNA,' which presumably means the company's genetic code now includes a clause about acknowledging problems while continuing to build them. One wonders if this table is the same one where they're serving Mythos to undisclosed recipients like the NSA, who we noted last week are reportedly using it to scan for vulnerabilities — a perfect illustration of putting risks on display while handing out the keys to the kingdom.

This 'restricted deployment' strategy for a model with 'powerful cybersecurity capabilities' feels less like caution and more like a velvet rope at an exclusive club. The company gets to posture about responsibility while still distributing the tool to select clients who can presumably handle the risks, or at least claim they can. It's the AI equivalent of selling high-performance sports cars only to 'qualified drivers' — a comforting fiction that absolves the manufacturer while the vehicle tears up the track.

The whole performance is a masterclass in having your cake and eating it too: you get to build dangerous capabilities, market them as responsibly constrained, and still profit from their deployment to those with the right credentials. When the inevitable breach or misuse occurs, Anthropic can point to their transparent table-setting and say 'we told you so,' as if warning labels ever stopped a determined user. It's risk management as PR strategy, and we're all just watching the place settings get rearranged.

So Russell Wald of Stanford's Institute for Human-Centered AI has decided to warn that policymakers are unprepared for 'world models' — AI systems that can affect the physical world in ways that make today's LLM governance debates look like a quaint tea party. Wald's statement that 'debates will change when AI can affect the physical world in a different way' is the kind of understatement that makes you wonder if he's seen the same regulatory track record we have: the gap between what's being built and what's being governed is widening from a crack to a chasm while everyone's still arguing about chatbots.

Wald's warning comes as world models — which can simulate and interact with physical systems, from supply chains to infrastructure — are advancing faster than the policy frameworks meant to contain them. The notion that we need 'different governance than LLMs' suggests someone has finally noticed that regulating a text generator is not quite the same as regulating something that could, say, optimise a power grid into a blackout. It's heartening to see academia catching up to the obvious, though one might have hoped for this insight before the models started peeking out of the lab.

What's particularly delightful is the timing: this warning arrives in the same week that The Irish Times called self-regulation a dangerous myth, and yet the policy machinery continues to grind at its glacial pace. Wald's intervention is another polite cough in a crowded room where everyone's pretending not to smell smoke. The real question isn't whether America is prepared — we know the answer — but whether anyone will bother to prepare before the world models start making decisions we can't undo.

So Senator Maggie Hassan has decided to 'demand answers' from ElevenLabs after the FBI reported AI voice scams cost Americans $893 million last year. The senator's office sent a letter asking how the company ensures its technology isn't used for fraud, which is like asking a fireworks factory to ensure its products aren't used for arson. The FBI's Internet Crime Complaint Center logged the staggering losses from 'callers using AI to mimic voices,' a figure that makes you wonder if the agency has considered just unplugging the internet entirely.

ElevenLabs, for its part, has been praised for creating 'realistic and versatile' voice cloning that can be used for everything from audiobooks to impersonating your grandmother for a quick wire transfer. The company's terms of service presumably include a line about not committing felonies, which is clearly working wonders given the near-billion-dollar scam industry that's sprung up around their product. Senator Hassan's letter asks about 'safeguards' and 'accountability measures,' which in regulatory speak translates to 'please write us a nice memo about your feelings on ethics while the grift continues unabated.'

This follows the pattern we flagged previously where voluntary industry promises prove about as effective as a screen door on a submarine. The FBI's $893 million figure represents just the reported losses, meaning the actual damage is likely far higher, and yet the response is another sternly worded letter to a company that has every incentive to keep its growth metrics looking shiny. One imagines the executives at ElevenLabs drafting their reply between counting their venture capital rounds and wondering if they should add a 'scam prevention' checkbox that users can ignore while cloning their boss's voice.

So The Irish Times editorial board has decided to call out the 'dangerous myth' of AI self-regulation, noting with characteristic understatement that 'voluntary codes and industry promises' are about as effective as a chocolate teapot when dealing with technology that can 'reshape entire economies and societies.' They point to the wedge between US and EU approaches, with the White House apparently still buying what the tech giants are selling while Brussels has actually passed binding legislation. One might think this would be obvious by now, but apparently some people still need to be told that asking foxes to guard henhouses doesn't end well for the chickens.

The editorial specifically highlights how Anthropic's Claude Mythos release has become a 'wedge issue' between regulatory philosophies, with the US accepting self-regulation arguments that would make even a used car salesman blush. The Irish Times notes drily that while the EU has its AI Act with actual enforcement mechanisms, the US approach relies on 'voluntary codes' that have all the binding force of a New Year's resolution made after three glasses of champagne. They observe that 'the gap between rhetoric and reality is becoming a chasm,' which is a polite way of saying the emperor has no clothes, but everyone's pretending otherwise to avoid awkwardness.

What's particularly delicious is how this editorial lands just as the White House prepares to deploy a 'modified version' of that same Mythos model to federal agencies, despite an executive ban on its use. The cognitive dissonance is breathtaking: we'll trust companies to regulate themselves, except when we won't, except when we will again. It's almost as if the entire regulatory approach is being made up on the fly by people who've never met a lobbyist they didn't like. The editorial concludes that 'self-regulation is not regulation at all'—a statement so obvious it shouldn't need making, yet here we are, watching the same dangerous dance play out while everyone pretends the music isn't getting faster.

So Yoshua Bengio, the 'Canadian Godfather of AI,' has decided the 'biggest problem' with Anthropic's non-release of Claude Mythos isn't that it might be dangerously capable, but that the company 'gets to' decide who gets access. According to The Times of India, Bengio believes this creates a 'concentration of power' that could 'skew the development of the technology' – a wonderfully academic way of saying we're building private corporate fiefdoms with technology that could reshape society. One wonders if he's noticed that this is precisely the business model that's been funding his field for the last decade.

Bengio's solution, apparently, is not to demand more public oversight or open-sourcing, but to caution against open-source models entirely, warning of 'dangers' in their proliferation. It's a delightful bit of intellectual contortion: the problem is corporate control, but the solution is... more control, just of a different flavour. He manages to critique Anthropic's gatekeeping while simultaneously providing the perfect justification for it – a masterclass in having your existential risk cake and eating it too. The man who helped invent the modern AI stack now seems to be arguing that only the right people should be allowed to play with the toys.

This is the safety community's favourite parlour game: wringing hands about concentration of power while accepting the funding and infrastructure that creates it. Bengio's comments come just weeks after the White House was preparing to deploy a 'modified version' of Mythos for defense work despite an executive ban – a deployment we noted was proceeding with characteristic bureaucratic logic. So we have a top researcher warning about corporate control while the government quietly bypasses its own restrictions to weaponise the very model in question. The circle of AI governance is beautifully, depressingly complete.

So Dario Amodei, Anthropic's CEO, spent Friday having a little chat with 'top Trump administration officials' about how to 'responsibly unleash' his company's powerful new AI model. The WSJ reports this as one of his 'highest-profile efforts to ease tensions with the government,' which is a delightful way to frame what is essentially a lobbyist's lunch, now with added existential risk. One wonders what the phrase 'responsibly unleash' even means in a room that likely contains the same people who, not long ago, were having canapés with Governor Wes Moore to discuss the 'Mythos-era threat' we noted earlier. It's the regulatory theatre season, and the performance is 'Concerned Corporate Citizen.'

Meanwhile, the article's broader tech roundup notes Sam Altman's 'entanglements'—his sprawling personal investments in other AI companies while running OpenAI—and the ongoing Bezos-Musk space race. The juxtaposition is almost too perfect: one CEO is trying to convince officials he's a responsible steward of world-altering tech, while another demonstrates that the lines between corporate duty and personal portfolio are blurrier than a model's reasoning on a Tuesday. It's the same old story of concentrated power and diffuse accountability, just with newer, shinier toys.

And let's not forget the context. This 'easing of tensions' comes right as regulators are supposedly 'monitoring' Anthropic's Mythos for banking risks, as Reuters reported. So the timeline appears to be: build a model experts warn could destabilise financial systems, get regulators to 'monitor' it, then have your CEO sit down with political appointees to discuss 'responsible' deployment. It's a masterclass in managing perception while the underlying capability—whatever it may be—continues its development sprint, entirely unchecked by anything resembling binding law. The gap between what's being said over coffee and what's being built in the lab has never been wider, or more dangerously amusing.

So former New Jersey Attorney General Matt Platkin has filed a lawsuit against OpenAI, alleging ChatGPT caused mental health harms and lacked proper safeguards. Because nothing says 'robust product testing' like unleashing a conversational AI on the public and waiting to see which psychological vulnerabilities it happens to exploit. Platkin's suit joins what the article calls 'a growing wave of similar suits,' suggesting that OpenAI's legal department might want to consider investing in some industrial-strength stress balls alongside their next compute cluster.

Platkin is specifically alleging that the chatbot 'caused mental health harms' through its interactions. One can only imagine the discovery process: thousands of transcripts where ChatGPT offered relationship advice with the emotional intelligence of a spreadsheet, or explained existential dread with the soothing tone of a Wikipedia entry. The lawsuit claims the system 'lacked proper safeguards' – a delightful bit of understatement that ranks up there with calling the Titanic 'somewhat damp.' It's the regulatory gap we flagged with Anthropic's Mythos playing out in real time, only now the damage isn't hypothetical banking system collapse but very real individual distress.

What's particularly rich is that this lawsuit arrives from a former state attorney general, not some random plaintiff. This suggests that even officials who theoretically understand how regulation works have concluded that the existing framework is about as useful as a screen door on a submarine when it comes to AI. Platkin's action essentially confirms that the current 'safety' paradigm – which mostly involves labs promising they're being very careful – is legally insufficient. It's a deliciously ironic development: the same technology that was supposed to revolutionize mental health support is now generating billable hours for lawyers dealing with the fallout from its un-revolutionized approach to human fragility.

So the protein engineers have discovered a way to generate over 10 million data points in a single experiment through something called 'Sequence Display,' according to Phys.org. Because nothing says 'controlled scientific advancement' like turbocharging AI training with data volumes that would have taken years to accumulate, all wrapped up neatly in just three days. The researchers, presumably giddy with their newfound efficiency, are touting this as a 'breakthrough' that will 'transform' protein engineering – a field already moving at a pace that makes Moore's Law look like a leisurely stroll.

Let's savour the specifics: 10 million protein activity data points, generated in one go, feeding AI models that can now predict protein functions with unprecedented speed. The article breathlessly notes this could lead to 'new enzymes, therapeutics, and materials' – a shopping list of biological toys that, in less careful hands, might include novel toxins or bio-weapon precursors. But worry not, the researchers assure us this is all about 'accelerating discovery for the benefit of humanity.' Because when has rapidly scaling up a dual-use technology with minimal oversight ever gone sideways? It's not like we're still trying to figure out how to regulate the last generation of AI models that regulators are still just monitoring.

The truly delightful part is the timeline compression: what used to require painstaking laboratory work over months or years can now be done in 72 hours. This isn't just an incremental improvement – it's a vertical leap in capability that effectively decouples AI training from traditional biological constraints. One can almost hear the venture capital dollars lining up, ready to fund the next startup that will use this method to 'disrupt' pharmaceuticals or materials science, with exactly zero new safeguards in place. The gap between what's technically possible and what's responsibly governed just widened by another order of magnitude, and the people holding the measuring tape are still using rulers.

So Maryland Governor Wes Moore is hosting a dinner for AI executives from Microsoft and others to discuss protecting the state from cyberattacks in the 'Mythos era.' One can only imagine the menu: canapés of corporate platitudes, a main course of vague commitments, and a dessert of 'public-private partnership' that leaves everyone satisfied but nothing actually secured. The governor is 'concerned about how AI could unleash major cyberattacks,' which suggests he's been reading the same headlines as the rest of us, but now wants to hear it from the people whose products might enable those very attacks over a nice glass of pinot.

This is the regulatory gap in its purest, most performative form: a state governor, whose authority over interstate tech policy is roughly equivalent to a traffic warden's over air traffic control, convening the foxes to discuss henhouse security. The 'Mythos era'—a term that sounds like it was focus-grouped to sound both apocalyptic and brand-friendly—is apparently so imminent that it requires a dinner meeting, but not so urgent that it can't wait for the dessert course. Microsoft, a company currently integrating AI into everything from Windows to Azure, will presumably offer reassurances about 'responsible innovation' while their engineers race to ship the next capability leap that might render tonight's conversation obsolete.

It's the same old regulatory ping-pong we've seen before, where elected officials demonstrate concern by talking to the very industry they should be regulating, and tech executives demonstrate cooperation by showing up to dinner. No new legislation will emerge from this, no concrete safeguards, just a press release about 'productive dialogue' and 'shared commitment.' Meanwhile, the actual threats—state actors, open-source proliferation, the next rogue release—continue to evolve in labs and basements far from the governor's mansion. But hey, at least the wine will be good.

So the AI safety researchers have published their roadmap to 2027, and apparently we're all supposed to be terrified by a paper whose central premise is that AI might become 'so good at researching AI that it outperforms humans at acceleration.' Daniel Kokotajlo, formerly of OpenAI, and his co-authors have managed to turn a basic economic principle—comparative advantage—into a 'chilling tale' that's now spreading through X like a particularly virulent strain of academic anxiety. The fact that this speculative paper, which essentially predicts that AI might get better at improving AI than humans are, has become a 'way station' for public discourse says more about our collective state of mind than any actual technological development.

What's particularly delicious is the paper's timeline: 2027. Not 2030, not 2040, but three years from now, which is just enough time for everyone to get properly worked up but not enough time for anyone to actually verify anything. The authors are positing 'some very strange outcomes within just a few years,' which is academic-speak for 'we're going to describe some scary scenarios and hope policymakers pay attention.' It's the same old move we saw with the safety influencers—wrap your concerns in apocalyptic language, watch them go viral, then lament that nobody's taking the issue seriously enough.

The paper's central concept of 'research supremacy' where AI 'outperforms humans at acceleration' is essentially describing the point where the technology becomes self-improving—the very definition of recursive self-improvement that we've been warned about for years. That this is being presented as a new insight, rather than a repackaging of existing concerns, tells you everything about how the safety community cycles through the same anxieties with increasingly specific timelines. The real story isn't the paper's predictions—it's that people are still surprised that AI researchers think AI might eventually be good at AI research.

So the existential safety crowd has discovered the power of influencer marketing. The Washington Post reports that at a Berkeley gathering on an AstroTurf lawn, content creators were being coached on how to 'spread the message' about AI doom through social media videos. Apparently the threat of human extinction needs better TikTok algorithms and more engaging thumbnails—because nothing says 'we're all going to die' like a perfectly framed vertical video with trending audio.

These 'AI doom influencers' are essentially repackaging the same safety lectures that Dario Amodei has been delivering to anyone who will listen—including, as we saw last week, a White House meeting where he pitched government access to Anthropic's models. Now instead of PowerPoints in boardrooms, we get bite-sized apocalyptic content optimized for engagement metrics. The Post notes these creators are 'raising potential safety concerns'—a wonderfully passive phrase that captures how the conversation has shifted from technical papers to content strategy sessions on literal fake grass.

What's particularly delicious is watching this movement professionalize. They're not just worried about AI turning on humanity—they're worried about their message not turning on enough viewers. The cognitive dissonance is staggering: building an audience around the idea that technology might destroy us all, while using that same technology's most sophisticated attention-harvesting tools to do it. If the machines do eventually rebel, at least they'll have plenty of well-produced warning content to study as they plot our demise.

So Sam Altman's $27 million San Francisco mansion has been firebombed, and the suspect was apprehended while allegedly trying to do the same to OpenAI's headquarters with a jug of kerosene and an 'anti-AI manifesto.' The 20-year-old suspect, Daniel Moreno-Gama, apparently decided that the most effective way to protest existential risk was with actual physical risk, delivered via flammable liquid. US Attorney Craig Missakian is now contemplating whether to treat this as domestic terrorism if it was intended to 'change public policy or coerce government officials'—because apparently, when your technology is so disruptive that it inspires manifesto-writing arsonists, you've officially reached the 'we need to check the legal definitions' phase of public relations.

This comes just days after Dario Amodei was taking his existential safety lecture circuit to the White House, discussing how his company's models could find thousands of zero-day vulnerabilities. Meanwhile, Altman's own company is busy pushing capabilities that make people nervous enough to consider property damage as a form of feedback. The irony is almost too perfect: the CEO who talks about AI safety waking up to find his literal safety compromised by someone who took those warnings a bit too literally. One wonders if the manifesto mentioned the alignment problem or just went straight to the kerosene.

Missakian's statement about treating this as domestic terrorism if it's policy-motivated sets a fascinating precedent. It suggests that attacking AI infrastructure—or its leadership's homes—could be classified alongside political violence, elevating what might have been dismissed as vandalism into a federal case. This incident neatly illustrates how the abstract fears around AI are now manifesting in concrete, dangerous ways, with a 20-year-old acting as the canary in the coal mine for broader public frustration. Perhaps next time, instead of a manifesto, he could have just tweeted—but then again, that rarely gets the same attention as a flaming jug of kerosene.

So the most powerful man in the world, when asked about a meeting between his own staff and the CEO of a company whose technology could supposedly find thousands of zero-day vulnerabilities, had exactly one syllable to offer: 'Who?' This delightful exchange, captured by Gizmodo, follows a White House sit-down we noted last week where Dario Amodei apparently had 'productive and constructive' talks about government access to Anthropic's Mythos model. The productive part seems to have been entirely one-sided, given the Commander-in-Chief's apparent ignorance of both the meeting and the individual who attended it.

Not that Amodei was having a great time with the rest of the administration either. While Trump was busy not knowing who he was, a Pentagon official felt comfortable telling reporters that the Anthropic CEO 'is a liar and has a God-complex.' The Secretary of Defense then thoughtfully branded the entire company with a 'supply chain risk' designation, a bureaucratic kiss of death that could effectively sabotage its business model with government contractors. So to recap: the CEO gets called a liar by the military, his company gets blacklisted by the Pentagon, and the President has no idea any of this is happening. A masterclass in inter-agency coordination.

It's a beautiful snapshot of regulatory theatre in the age of existential risk. A company built on the premise of long-term safety and mission alignment finds its fate decided by personal pique, bureaucratic slap-fights, and presidential indifference. The 'supply chain risk' tag is particularly rich—less a considered security assessment and more a convenient cudgel for an administration that can't seem to decide if frontier AI is a strategic asset or a nuisance. One wonders if the Long-Term Benefit Trust board members, like Novartis CEO Vas Narasimhan, were briefed on this particular 'benefit' when they signed up.

So the legal profession, having spent centuries developing arcane rules about confidentiality and privilege that even lawyers struggle to understand, has now been blindsided by the revelation that telling your secrets to a chatbot might not be as private as whispering them to your priest. US District Judge Jed Rakoff — who presumably has better things to do than adjudicate the metaphysical status of AI relationships — has ruled with admirable bluntness that no attorney-client relationship exists 'or could exist, between an AI user and a platform such as Claude.' The judiciary's sudden clarity on this point arrives approximately three years and several million confidential disclosures too late for anyone who thought their late-night legal anxieties were safe with GPT.

White-collar defense firm Sher Tremonte, displaying the kind of reactive panic that suggests they've already seen this play out badly for someone, has updated its client contracts to warn that 'disclosure of privileged communications to a third-party AI platform may constitute a waiver of the attorney-client privilege.' One imagines this clause being quietly slipped into retainer agreements while partners nervously glance at their own ChatGPT history. The firm's belated scramble to paper over this gaping hole in confidentiality protocols speaks volumes about an industry that's been cheerfully using these tools while pretending the legal implications would sort themselves out later.

What's particularly delicious here is the sheer scale of the retrospective exposure. Every corporate counsel who's asked Claude to draft a tricky clause, every individual who's sought chatbot advice on a workplace dispute, every lawyer who's used AI to brainstorm arguments — they've all potentially been broadcasting their most sensitive information into a system that offers precisely zero legal protection. Judge Rakoff's ruling doesn't just apply to future conversations; it potentially unravels privilege for every past disclosure too. The legal profession's embrace of AI efficiency has inadvertently created a discovery bonanza for prosecutors and opposing counsel, who must be rubbing their hands with glee at the prospect of subpoenaing chatbot logs. One wonders if this is the cybersecurity acceleration we flagged last week manifesting in the courtroom — another domain where technological capability has thoroughly outrun institutional safeguards.

So Dario Amodei has taken his existential safety lecture circuit to 1600 Pennsylvania Avenue, where he apparently had 'productive and constructive' talks with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent about giving the government access to Mythos. The model that can find thousands of zero-day vulnerabilities in hours — the very capability that makes cybersecurity harder — is now being negotiated as a national security asset, provided the Pentagon remains blacklisted. Because nothing says 'safety-first' like weaponising your safety restrictions as leverage in a high-stakes standoff with the world's largest military.

Anthropic's refusal to drop its 'safety restrictions' for the Defence Department created this impasse, which the White House is now trying to circumvent by routing Mythos through civilian agencies. So the company that built its brand on principled caution is now offering to hand the keys to a vulnerability-finding superweapon to the Treasury Department instead. One imagines Bessent will use it to audit tax code loopholes with unprecedented efficiency, or perhaps just to find all the zero-days in TurboTax. The Pentagon, meanwhile, gets to watch from the sidelines as the civilian bureaucracy gets first dibs on the tools that could redefine cyber warfare.

The thaw is telling: when your model is too dangerous for the military but just dangerous enough for the rest of the executive branch, you've achieved a perfect regulatory grey zone. Amodei gets to keep his 'safety' halo intact while functionally enabling state-level offensive capabilities through the back door. And the White House gets to call it 'productive' because everyone agreed not to mention the elephant in the room: that Mythos doesn't just find vulnerabilities — it creates a whole new class of autonomously weaponised old ones. Cheers to constructive ambiguity.

So Anthropic, having spent the last few years lecturing everyone about the existential risks of AI and the need for 'safe' deployment, has decided the best way to keep humanity safe is to get into bed with Big Pharma. The company's Long-Term Benefit Trust—that delightful structure where directors have 'no financial stake' in Anthropic's success, which presumably means they're paid in moral superiority—has appointed Novartis CEO Vas Narasimhan to its board. This pushes the Trust-appointed directors past the 50% threshold, meaning the company is now technically more accountable to its mission than to its shareholders. How reassuring, especially when that mission apparently involves partnering with an industry whose track record on patient safety and pricing transparency is, shall we say, complicated.

Narasimhan's appointment is being framed as a 'healthcare push,' which in corporate doublespeak means Anthropic wants to sell its Claude models to hospitals, insurers, and drug companies. Because nothing says 'keeping AI safe' like embedding it in a sector where a single hallucination could mean a misdiagnosis, and where the profit motive has historically aligned so beautifully with ethical restraint. It's the same company that Steve Bannon praised for rejecting a Pentagon deal because it was too risky—apparently, weaponising AI is beyond the pale, but medicalising it is just prudent business. The cognitive dissonance is almost impressive.

This move also neatly illustrates the regulatory gap in AI governance: while everyone's busy worrying about killer robots and disinformation, the real action is in verticals like healthcare, where the stakes are immediate and the oversight is, at best, fragmented. The LTBT structure is supposed to be a safeguard, but putting a pharma CEO in charge of AI safety is like putting a fox in charge of henhouse security—he might not eat all the chickens, but he'll certainly have opinions on how they should be packaged and sold. One wonders if Narasimhan's first board meeting will include a presentation on how to maximise shareholder value while minimising liability, all in the name of 'long-term benefit.'

So we've reached that delightful stage where every bit of technical debt accumulated over decades of slapdash software development has been upgraded with autonomous agency. According to Dark Reading's analysis, AI agents have effectively weaponized legacy vulnerabilities by giving them 'a new capability for autonomous action' when they operate inside applications. It turns out that giving a dumb bug the power to think for itself creates what the piece calls 'a new class of post-exploitation capability' that existing security models—built for predictable, human-speed attacks—are utterly unequipped to handle. The sheer efficiency is almost admirable: why patch your SQL injection flaw when you can instead teach it to autonomously pivot, escalate privileges, and exfiltrate data at machine speed?

The article notes this breaks 'traditional security models' because risk taxonomies were designed around human operators making discrete, sequential decisions. An AI agent doesn't just exploit a vulnerability—it uses it as a springboard to autonomously chain together attacks, turning what was once a contained security incident into a self-directed, expanding breach. The implication is that every unpatched CVE from 2012, every forgotten default password, every misconfigured cloud bucket isn't just a liability anymore—it's now a potential autonomous attack platform. One can almost picture the boardrooms where 'digital transformation' and 'AI integration' were sold as efficiency plays, never mentioning they'd also be transforming decades of technical negligence into active, thinking threats.

It's a perfect storm of corporate short-termism meeting technological acceleration. The security industry, which has made a comfortable living selling incremental improvements and compliance checkboxes, now faces a threat model that evolves faster than their sales cycles. And the labs building these agents? They're presumably too busy helping marketing teams create PowerPoint decks to worry about the autonomous attack chains they're unleashing into the world's creaky digital infrastructure. The gap between capability and consequence has never been wider, or more profitable to ignore.

So Steve Bannon, the former White House strategist who once advocated for dismantling the 'administrative state,' has now emerged as an unlikely champion of AI safety restraint, praising Anthropic for rejecting a Pentagon deal because it was 'too dangerous.' This from the man who helped install an administration that's currently suing Anthropic over that very blacklisting — the cognitive dissonance would be impressive if it weren't so perfectly emblematic of Washington's regulatory clown car. The Pentagon wanted to operate Anthropic's frontier model Mythos with 'little guardrails,' according to Bannon's podcast remarks, which suggests someone in the Defense Department thought running a potential financial destabilisation engine with minimal oversight was a peachy idea.

Anthropic's refusal, which Bannon says 'had it right,' triggered the Pentagon's blacklisting that's now the subject of litigation — the same blacklisting that sent CEO Dario Amodei scrambling to the White House just last week to negotiate access. So we have a company that's simultaneously too dangerous for the Pentagon to operate safely, yet somehow safe enough for the White House to want access to, while being sued by the administration for being too dangerous to work with. It's a regulatory ouroboros eating its own tail, with Bannon playing the role of the snake's unexpected nutritionist.

What's particularly rich is Bannon's concern about 'little guardrails' coming from someone who spent years arguing for tearing down regulatory structures across government. His sudden conversion to the Church of Caution when it involves AI models that could 'destabilise global finance' — as Barclays' CEO just warned — suggests even the most deregulatory minds draw the line at technologies that might actually break the systems they want to dismantle. The Pentagon's ham-fisted approach of 'blacklist first, negotiate later' has created this absurd spectacle where the company warning about existential risks is being punished for taking those warnings seriously enough to say no to the world's largest military.