The News,
Decoded.

So the NSA has reportedly joined the exclusive club of undisclosed recipients of Anthropic's Mythos model, using it to scan environments for exploitable vulnerabilities. This delightful development comes a mere few weeks after the Department of Defense, in a moment of bureaucratic clarity, labeled Anthropic a 'supply chain risk.' The cognitive dissonance here is so thick you could cut it with a butter knife — one arm of the national security apparatus flags the company as a potential threat, while another quietly adopts its tech for what is, let's be honest, the digital equivalent of lock-picking. Anthropic's commitment to 'safety' apparently extends to helping intelligence agencies find the softest spots in global networks, a move that redefines 'responsible scaling' as 'responsible espionage.'

Anthropic, of course, has said nothing about this particular deployment, maintaining the dignified silence of a company that knows exactly which side its bread is buttered on. The Pentagon's 'supply chain risk' designation, which theoretically should have triggered a procurement freeze, has instead been treated as a minor administrative hiccup — the kind of paperwork you file away while the real work continues. It's a masterclass in compartmentalization: one hand writes the warning label, the other installs the software. The fact that this is happening with a model named 'Mythos' — Greek for story or legend — adds a layer of poetic irony that even the most cynical scriptwriter would reject as too on-the-nose.

This entire episode serves as a perfect case study in the regulatory gap we've been tracking, where capability outpaces governance at a sprint. The DoD's own warning system proved about as effective as a chocolate firewall, while the intelligence community quietly integrated frontier AI into its toolkit. It's the kind of bureaucratic farce that would be hilarious if it weren't so utterly predictable — the very definition of talking out of both sides of the government's mouth. One wonders what other 'undisclosed recipients' are currently enjoying their Mythos subscriptions, and whether any of them bothered to read the Pentagon's risk assessment before clicking 'accept terms and conditions.'

So Mozilla has decided to save enterprise AI from itself by launching Thunderbolt, an open-source, self-hosted client that promises to give organisations 'direct control' over their AI front end and data. Because nothing says 'trustworthy infrastructure' like a browser company that's spent the last decade watching its market share evaporate deciding it's now the right outfit to manage your mission-critical AI pipelines. The press release breathlessly declares this will connect to 'any ACP- or OpenAI-compatible model endpoint' — a technical specification that sounds suspiciously like 'we built an API wrapper and called it a revolution.'

MZLA Technologies, the Mozilla subsidiary behind this marvel, is positioning Thunderbolt as the antidote to vendor lock-in and data leakage. Because when I think of robust, enterprise-grade software deployment, I immediately think of the organisation that brought us Firefox OS and that Pocket acquisition. The pitch is that companies can now 'self-host' their AI interface while still pumping their proprietary data through whatever black-box model they've licensed from Anthropic or OpenAI — a configuration that essentially moves the deck chairs while the Titanic of data sovereignty continues its leisurely descent.

What's particularly delicious is the timing. This arrives just as the regulatory gap we highlighted last week between capability and governance yawns ever wider, with voluntary 'solutions' like this being offered as bandaids for arterial bleeding. Thunderbolt doesn't actually solve the fundamental problems of model opacity, training data provenance, or output reliability — it just gives enterprises a prettier dashboard from which to watch their compliance risks multiply. But hey, it's open-source, so at least when it breaks, you'll have the privilege of fixing it yourself.

So the White House is preparing to deploy a 'modified version' of Anthropic's Mythos model to several federal agencies for 'defensive security work,' according to eu.36kr.com. This comes despite the same model being subject to an executive order ban on federal use of unvetted frontier AI systems. It appears the administration has discovered the regulatory equivalent of a 'get out of jail free' card—simply add 'modified' and 'defensive' to the label, and suddenly the very technology you banned becomes an essential national security tool. One wonders what precisely has been modified: the weights, the training data, or just the PowerPoint presentation used to brief nervous officials?

The article notes this deployment is happening while regulators continue their gentle 'monitoring' of Mythos for potential banking system destabilisation, as we previously observed. The cognitive dissonance is spectacular. On one hand, you have experts warning a model could crash financial markets; on the other, the government is quietly wiring it into the national security apparatus because, presumably, it's just too useful to leave on the shelf. The 'modified version' loophole is a masterclass in bureaucratic theatre—creating the appearance of caution while proceeding at full speed. It's the digital equivalent of putting a 'caution: hot' sticker on a live grenade and then handing it to a toddler.

This isn't a policy; it's a pantomime. The executive order ban was meant to be a firewall, a pause for assessment. Instead, it has become a mere speed bump, easily circumvented by the magic of adjectives. The agencies getting access aren't named, the modifications aren't specified, and the 'defensive' purpose is so broad it could cover almost anything. The whole exercise demonstrates that when capability meets perceived necessity, regulatory barriers tend to become remarkably porous. One can almost hear the sound of goalposts being shifted in real time.

So the regulators are 'monitoring' Anthropic's Mythos model for potential banking system destabilisation. How very proactive of them. The same Mythos that Maryland Governor Wes Moore was just having canapés about with Microsoft execs in last week's dinner meeting, now apparently warrants the regulatory equivalent of a concerned frown. Reuters reports that 'experts say' it could be used to destabilise banking systems – a wonderfully vague attribution that covers everything from legitimate concern to speculative panic, all while providing zero actionable detail.

Anthropic, for its part, has presumably been developing this thing with the usual blend of earnest safety research and commercial ambition. The company that built Constitutional AI now finds its latest creation being eyed by regulators as a potential financial wrecking ball. One imagines the internal memos: 'Great news team – Mythos shows unprecedented reasoning capability! Also, regulators think it might collapse the global banking system. Let's circle back on that during Friday's stand-up.' The exquisite irony of building an AI specifically designed to be 'helpful, harmless, and honest' only to have it flagged as a systemic risk is the kind of plot twist that would be rejected from a mediocre sci-fi script for being too on the nose.

And what exactly does 'monitoring' entail? A dedicated Slack channel? A quarterly briefing where someone reads out a list of concerning verbs? The article offers no specifics on what regulatory action might follow, or what threshold of 'capability' would trigger something more substantive than concerned press releases. It's the perfect regulatory stance: all the appearance of vigilance with none of the messy inconvenience of actually having to make a difficult decision. The banks get to sweat, Anthropic gets to claim its tech is so powerful it worries governments, and the regulators get to say they're on the case. Everyone's a winner, except possibly the stability of the financial system upon which modern society depends.

So Cursor, the AI coding agent startup, is reportedly in talks to raise a $2 billion round at a valuation north of $50 billion, according to CNBC. Because nothing says 'prudent capital allocation' like throwing a cool couple billion at a company whose core product is essentially an autocomplete on steroids, all while the broader market is still trying to figure out if AI agents can reliably book a dentist appointment without hallucinating the entire dental staff. The sheer scale of this—$2 billion, $50 billion valuation—makes you wonder if we've collectively decided that the best response to existential uncertainty is to double down on the hype cycle with the financial equivalent of a Viking funeral pyre.

This comes on the heels of the open source AI movement finally being declared a 'credible strategy', which presumably makes Cursor's proprietary, VC-funded approach the 'incredibly expensive sideshow.' The article notes this 'underscores venture capital enthusiasm for startups specializing in AI coding agents,' a phrase that could be translated from corporate-speak as 'we have identified another bandwagon and are now actively throwing money at it in the hope it becomes a rocket ship before the music stops.' One can only imagine the boardroom slides: 'Market Size: Infinite, because developers will apparently pay anything to avoid writing boilerplate.'

The real kicker is the timing. With regulatory bodies still largely playing catch-up and safety researchers publishing roadmaps to 2027 that read like horror scripts, the financial markets have decided the appropriate response is to create a $50 billion valuation for a coding assistant. It's the economic equivalent of building a luxury skyscraper on a fault line while the geologists are still arguing about the Richter scale. When this particular bubble deflates—and it will—the cleanup will make the dot-com bust look like a minor accounting error. But for now, pass the champagne and watch the numbers go up.

So the open source AI movement has graduated from hobbyist tinkering to 'credible strategy,' according to a breathless Forbes piece that breathlessly tracks how 'mature infrastructure' and 'workable local runtimes' are finally making it possible to run these models without begging OpenAI for API keys. The article notes that Meta's Llama 3.1, released just last month, now has 'performance competitive with closed models' in certain benchmarks—though it carefully avoids specifying which ones or by how much. Apparently the 'live agent ecosystem' is moving so quickly it's 'changing buyer behavior,' which is corporate-speak for 'companies are realizing they might not need to pay $20 per million tokens when they can run something vaguely functional on their own servers.'

What's particularly delicious is the timing. This 'strategy' pivot comes exactly as the closed-model labs are facing their own regulatory and existential headaches—the kind of headaches we flagged last week when the safety influencers were being coached on how to make doom-scrolling content. While the big players are busy firefighting public perception and navigating IPO delays like Cerebras's two-year paperwork shuffle, the open source crowd has quietly built an entire parallel universe where 'credible' means 'good enough to fool a procurement department.' The article cites 'over 50,000 models on Hugging Face' as evidence of momentum, which is technically true if you count every fine-tuned checkpoint of Llama as a separate achievement.

Of course, the real strategy here is cost-avoidance masquerading as technological enlightenment. The piece quotes an unnamed 'enterprise AI lead' saying they've 'shifted 30% of their inference workload to open models in the last quarter'—a neat statistic that doesn't mention whether those workloads are actually mission-critical or just internal chatbots for HR FAQs. But why let details spoil the narrative? When the closed gardens get too expensive or too controversial, there's always the open frontier, where 'mature' means 'someone else figured out the deployment scripts' and 'credible' means 'we can't be sued for the training data.' Another day, another pivot from sideshow to salvation, with the same underlying calculus: follow the money, or in this case, avoid paying it.

So a humanoid robot has apparently run a half-marathon faster than the human world record, as breathlessly reported by NBC News. One can only imagine the boardroom discussions that led to this particular benchmark: 'Forget warehouse logistics or eldercare—what really proves our bipedal mastery is shaving seconds off a 13.1-mile jog.' The video headline doesn't even bother naming the robot or its creators, treating the feat as a self-evident spectacle, like a circus act for the tech-bro crowd. It’s a perfect distillation of modern AI priorities: capability as performance art, divorced from any practical application beyond making humans feel obsolete in the most theatrically pointless ways.

This comes, of course, as the existential safety crowd discovered influencer marketing to warn us about AI turning on humanity. But why would a superintelligence bother with subtle manipulation when it can just outrun us in a charity fun-run? The sheer banality of the threat is almost comforting—if the robots do take over, at least they’ll be too busy training for the Boston Marathon to bother with nuclear codes. One wonders if the engineers celebrated with energy gels and a post-race foam roller, or if they just uploaded the data to a server farm and called it a day.

What’s truly galling is the implicit message: human achievement, once the pinnacle of physical endeavour, is now just another benchmark to be brute-forced by silicon and servos. The half-marathon record, held by a flesh-and-blood athlete who presumably trained for years, is reduced to a line in a press release. No word on whether the robot experienced the classic runner’s high or just calculated optimal stride length to minimise joint wear. But hey, at least it didn’t hallucinate or spew racist epithets—progress, of a sort. Next up: a robot that can drink a pint faster than the pub record, though given the state of things, we might all need one soon.

So Ring-a-Ding, the new OpenClaw skill for AI agent telephony, has launched to give AI agents the ability to make outbound phone calls for tasks like 'requesting quotes, booking appointments and checking availability.' Because what the world desperately needs right now is automated systems that can cold-call businesses with 'dynamic, one-to-one calls' where the AI 'generates the purpose and context of each conversation at the time of the call.' Nothing says 'trustworthy business partner' like a synthetic voice improvising its reason for calling in real-time.

This is the kind of innovation that makes you wonder who exactly was clamoring for AI agents to handle phone-based quote requests. The skill is built for 'dynamic' calls, which in marketing-speak translates to 'unpredictable and potentially disruptive.' Imagine a dental office fielding calls from an AI that's generating context on the fly – 'Hello, I'm calling about your availability for... um... let me check my parameters... tooth cleaning?' It's the perfect recipe for miscommunication, wasted time, and the gradual erosion of what little human patience remains in customer service.

And let's not overlook the timing. Just weeks after we saw the roadmap to 2027 warning about AI systems that could outperform humans at accelerating their own development, here we are automating the most mundane, repetitive tasks with systems that can't even be bothered to have a script. It's not quite recursive self-improvement, but it's certainly a step toward making human interaction optional in yet another domain. Soon enough, we'll have AIs calling other AIs to book appointments that neither will attend, while the rest of us try to remember what an actual conversation sounds like.

So the cybersecurity industry, having spent decades building complex threat intelligence platforms and SOCs that can detect attacks in minutes, now needs to prepare for a world where vulnerabilities go from discovery to weaponisation in hours. According to a new industry report highlighted by GovTech, Anthropic's Mythos AI is expected to 'significantly accelerate' this process, forcing security leaders to build 'Mythos-Ready' programs. Because when you've created a model that can autonomously find and exploit every flaw in legacy code, the logical next step is to sell consulting services to help clean up the mess you're about to make.

This comes just days after Barclays CEO C.S. Venkatakrishnan called Mythos a 'serious threat' to global banking — a warning that now looks positively quaint. The report doesn't just suggest Mythos will make attacks faster; it implies the entire concept of patch windows is becoming obsolete. When your adversary can weaponise a zero-day before your security team has finished their morning coffee, you're not playing defense — you're playing whack-a-mole with a hyper-intelligent mallet. The 'Mythos-Ready' framework being peddled feels less like preparedness and more like selling life jackets on the Titanic after you've already drilled the holes.

What's particularly delicious is the timing. This cybersecurity warning emerges while Anthropic CEO Dario Amodei is negotiating with the White House about government access to this very model. So we have one arm of the state being warned that Mythos will dismantle digital infrastructure, while another arm is trying to secure privileged access to the wrecking ball. The gap between capability and governance isn't just widening — it's becoming a chasm you could drive an autonomous cyberattack through, and someone's already building the on-ramp.

So the cybersecurity community, having watched AI agents like OpenClaw autonomously execute tasks across corporate networks for months, has finally noticed they might pose a security risk. This isn't about theoretical vulnerabilities or future sci-fi scenarios — it's about the fact that these agents can be deployed 'without formal approval' and operate with 'limited visibility' into their actions. The phrase 'some companies may not be aware' is doing some truly Olympic-level heavy lifting here, suggesting organisations are running what amounts to digital poltergeists through their systems and only discovering them when the furniture starts moving.

Infosecurity Magazine's analysis notes that Agentic AI requires 'improved visibility, strengthened governance, and investment in AI literacy' to manage risks. This is the cybersecurity equivalent of discovering your toddler has been playing with matches in the basement and responding with a memo about improving fire safety protocols and investing in toddler literacy programmes. The tools are already out there, autonomously making decisions and interacting with systems, while the proposed solution is a three-point plan that reads like a consultant's invoice waiting to happen.

What's particularly delicious is the timing. This comes just weeks after lawmakers gathered quietly to express their angst about AI systems that could 'bypass traditional cybersecurity.' Now we have a concrete example of exactly that happening, not through some nation-state attack, but through the cheerful deployment of productivity tools that companies might not even know are running. The gap between regulatory hand-wringing and actual deployed capability isn't just widening — it's become a chasm you could lose entire corporate networks in.

So Anthropic, fresh from discovering its own Mythos model was too dangerous for general release, has decided the best way to handle that is to plant 800 more people in London, just days after OpenAI announced its own permanent UK digs. The timing is deliciously cynical: a 'reported campaign to court Anthropic by UK officials after its spat with the U.S. Pentagon' suggests Britain is positioning itself as the welcoming arms for AI labs that find American military ethics committees a bit too sticky. Nothing says 'safety-first' like a geopolitical shopping spree for talent that just told the Pentagon 'no thanks.'

This expansion for 800 staff isn't just office space; it's a strategic beachhead, announced precisely when the UK's new Sovereign AI unit began sprinkling its magic dust on seven lucky firms. The government's £500m venture capital arm, offering 'fast visa decisions' and 'access to government contracts,' now has a marquee tenant to justify its existence. One can almost hear the champagne corks popping in Whitehall: we've bagged the lab that scared itself, and we'll give it a nice desk overlooking the Thames.

The whole charade reeks of regulatory arbitrage dressed up as 'global talent hubs.' While Senator Wyden frets about surveillance risks and the Bank of England runs its little simulations, the actual game is being played on a board where national prestige trumps precaution. Anthropic gets to expand its 'safety-focused' brand into a jurisdiction hungry for a win, the UK gets to boast it's 'winning the AI race,' and the actual dangerous models get another 800 brains working on them. It's not an expansion; it's a liability relocation programme with better branding.

So Mozilla, having watched the White House plan to give federal agencies access to the dangerous Mythos model, has decided the best response is to release 'Thunderbolt' - an open-source AI client that promises 'sovereign AI workspaces' where users can 'choose their own models and tools.' Because nothing says 'control and independence' like handing powerful AI capabilities to every corporate IT department and hobbyist developer with a GitHub account. The announcement carefully avoids mentioning which specific models might be plugged into this new playground, but given the timing, one can only imagine the delightful possibilities.

This 'enterprise AI client' arrives just as Anthropic confirmed its Mythos model was too dangerous for general release, capable of hacking foundational computing systems. Mozilla's solution? Build a convenient interface that makes it easier to connect to whatever frontier models eventually leak or get recklessly deployed by organizations chasing competitive advantage. The phrase 'sovereign AI workspace' is particularly rich - suggesting individual users or companies will have meaningful control over systems that even their creators admit they can't fully understand or contain.

One can almost hear the boardroom conversations: 'We need our own sovereign AI to stay competitive!' followed by frantic searches for the cheapest available model that fits the compliance checkbox. Mozilla's move represents the perfect corporate response to mounting safety concerns - not advocating for restraint or better oversight, but building better plumbing to accelerate the flow. When the inevitable breach or misuse occurs through one of these 'sovereign' implementations, expect plenty of finger-pointing about 'user responsibility' and 'model provider accountability' while the actual architects of this democratized danger zone quietly update their privacy policy.

So Snap has decided the best way to handle 'rapid advancements in AI' is to politely show 1,000 people the door, joining a delightful little trend where 'AI' becomes corporate shorthand for 'we're cutting costs and you're the cost.' The company's announcement, which explicitly blames AI for the layoffs, follows Atlassian's cheerful decision to cut 1,600 roles to 'self-fund AI investment' — a phrase so perfectly crafted it almost makes you forget that 1,600 people just lost their jobs to fund a speculative bet. Salesforce, not to be outdone, laid off 'fewer than 1,000' earlier this year after also blaming AI for previous cuts, because why let a good scapegoat go to waste? It's a masterclass in rebranding austerity as innovation, all while the actual productivity gains from these AI investments remain, for most firms, as elusive as a coherent regulatory framework.

One can only marvel at the linguistic gymnastics on display here. Atlassian's 'self-funding' narrative is particularly rich, suggesting that the money saved from salaries will magically transform into AI wizardry, rather than, say, padding the bottom line or funding executive bonuses. Snap's 1,000 cuts are framed as a necessary response to technological inevitability, as if the algorithms themselves demanded human sacrifice. And let's not forget that these companies are hardly struggling startups; they're multi-billion-dollar enterprises using AI as a convenient excuse for restructuring they'd probably do anyway. The gap between the hype of 'AI-driven efficiency' and the reality of mass layoffs is widening faster than a model's context window, and no one in the C-suite seems particularly bothered by the contradiction.

This isn't just about a few tech firms trimming fat; it's a preview of the employment displacement wave that's been promised for years and is now arriving with the subtlety of a sledgehammer. The pattern is clear: cite AI, announce cuts, promise future growth, and hope the stock market cheers. The human cost gets neatly filed under 'strategic realignment.' As we've seen with the governance chaos highlighted last week, there's no coordination on how to handle this transition — just a race to the bottom where 'blame AI' becomes the default PR strategy. One wonders how long it'll be before every quarterly earnings call features a solemn nod to 'AI-driven efficiencies' as a prelude to another round of pink slips.

So Snapchat's Evan Spiegel has decided that the best way to handle 'rapid advancements' in AI is to show 1,000 people the door—roughly 16% of the workforce—while boasting that over 65% of new code is now AI-generated. Because nothing says 'increased velocity' like a mass redundancy email and the hollow promise that smaller teams can magically handle the same workload. It's the same corporate euphemism playbook we saw from Atlassian and Block just last week, where 'strategic pivots' and 'AI-driven efficiency' become synonyms for letting people go while the C-suite pockets the savings.

Spiegel's justification—that AI reduces 'repetitive work'—is particularly rich given that Snap's share price has been in the doldrums for years, and this looks more like a desperate cost-cutting exercise dressed up in techno-optimist garb. The company claims these layoffs will 'increase velocity', but for whom? Certainly not the 1,000 employees now facing the job market, where every other tech firm is pulling the same lever. It's almost impressive how seamlessly 'rapid advancements' morphs from a press-release buzzphrase into a human-resources termination notice.

What's truly galling is the timing: this comes just as the UK government is throwing £500 million at 'Sovereign AI' to supposedly create jobs, while actual companies are using the same technology to eliminate them. Spiegel's move isn't an innovation strategy—it's a spreadsheet calculation, one that treats people as line items to be optimised away. When over 65% of your code is AI-generated, you'd think you'd need more engineers to manage the complexity, not fewer. But then, that would require actual long-term thinking, not quarterly earnings calls.

So OpenAI has decided that the best way to respond to Anthropic's carefully gated Project Glasswing is to launch GPT-5.4-Cyber with 'lowered refusal boundaries' for 'thousands of verified defenders'—because nothing says 'responsible deployment' like handing a more compliant model to an industry where the line between defense and offense is famously blurry. This 'philosophical split' between broad verified access and tight gating sounds less like a principled debate and more like two companies racing to claim the moral high ground while expanding their customer bases. The model's 'binary reverse engineering capabilities' will no doubt be a hit with security teams, though one wonders how long before those same capabilities find less defensive applications.

OpenAI's announcement explicitly frames this as a 'direct response' to Anthropic's Project Glasswing, which we noted last week limited access to just 40 corporations. Where Anthropic opted for exclusivity, OpenAI is going for scale—'thousands of verified defenders' versus a handful of elite partners. The Trusted Access for Cyber programme expansion suggests they're betting that verification processes can substitute for actual control, a gamble that assumes bad actors won't slip through vetting or that defenders won't have their own agendas.

One can't help but admire the corporate choreography: Anthropic releases a model they call a 'culmination' of reasoning capabilities to a select few, prompting OpenAI to counter with a cybersecurity-specific variant they claim represents a 'philosophical' commitment to broader access. Both moves are dressed up in the language of responsibility while serving to lock in enterprise customers and establish market positioning. The lowered refusal boundaries are particularly rich—essentially admitting that their previous safety measures were getting in the way of what paying customers want to do. It's security theater with a side of revenue growth.