The News,
Decoded.

So OpenAI, having decided that curing diseases is the perfect 'side quest' after all, has unveiled GPT-Rosalind, a frontier reasoning model for drug discovery. This comes just weeks after the company shed its OpenAI for Science initiative and the researcher who led it, Kevin Weil, in what was described as a focus on core priorities. Apparently 'focus' now means launching a new life sciences division with partners like Amgen and Moderna, proving once again that corporate strategy documents are written in disappearing ink.

The model is named after Rosalind Franklin, the crystallographer whose work was essential to understanding DNA's structure but who was famously sidelined from recognition. OpenAI's branding department clearly has a sense of irony, christening a tool built by a company that just sidelined its own science lead after a wave of researcher departures. They claim GPT-Rosalind can 'accelerate drug discovery' and is the 'first release' in a life sciences series, which sounds ambitious for an organization that recently decided such applied research was too distracting from the main event of building AGI.

This is classic lab behavior: announce a shiny new capability in a high-stakes field like medicine while quietly deprioritizing the foundational research that might make it safe or equitable. The partners—big pharma giants Amgen and Moderna—will get early access, naturally. So we have a black-box model from a company in regulatory crosshairs, aimed at biological systems, released amid a talent exodus from its safety and science teams. What could possibly go wrong in the hands of an industry known for its patient-first pricing strategies?

So OpenAI, having decided that customer-facing 'side quests' like Sora and its science initiative are too distracting from the main event, has bid farewell to two more senior researchers. Kevin Weil, who led OpenAI for Science, and Bill Peebles, the researcher behind the AI video tool Sora, both announced their departures on Friday, with Weil off to a new startup and Peebles returning to academia. It's a tidy cull of the non-core, a shedding of what CEO Sam Altman has reportedly deemed peripheral projects in favor of focusing on the frontier model race. Because when you're in a sprint to build the most powerful AI in history, who needs pesky distractions like commercializing your own video-generation marvel or applying your tech to scientific discovery? That's just clutter.

Weil's exit, in particular, feels like a quiet admission that the 'OpenAI for Science' initiative – launched with much fanfare about accelerating research – was less a moonshot and more a marketing footnote. The company is now 'recalibrating' its focus, which in corporate-speak means shutting down anything that doesn't directly feed the AGI altar. Peebles' departure, meanwhile, leaves Sora – a tool that generated genuine buzz – in a curious limbo, its lead architect walking away as OpenAI apparently decides that making money from video generation is a 'side quest' too far. It's a masterclass in strategic myopia: jettison the applied, monetizable innovations to double down on the theoretical, world-altering ones. Because what could possibly go wrong when you concentrate all your resources on a single, recursive goal?

This follows a pattern of high-profile exits at OpenAI, including co-founder Ilya Sutskever earlier this year, and echoes the broader industry trend where safety-focused researchers depart as companies prioritize capability leaps. The 'recalibration' here isn't just about trimming fat; it's a signal that OpenAI is narrowing its vision to a laser focus on frontier model development, potentially at the expense of safety checks and real-world applications. When the people who built your most public-facing tools are shown the door, it's not a pivot – it's a purge. And in the race to AGI, who has time for side quests when the main storyline is so all-consuming?

So OpenAI, fresh from planting its flag in London and watching Anthropic release a 'strategically safeguarded' Opus model, has decided the best way to demonstrate responsible AI development is to launch a series of models specifically designed for life sciences research. GPT-Rosalind, named after the scientist whose work was famously exploited without credit, is being pitched as a tool for 'fundamental reasoning' in biochemistry and genomics. Because what could possibly go wrong when you give AI models trained on biological data to researchers who are already warning about their potential misuse for designing dangerous pathogens?

Of course, the announcement comes with the usual corporate gloss: this is about 'helping life sciences researchers work faster,' not about creating the world's most accessible bioweapon design assistant. OpenAI's timing is impeccable, releasing this just as the safety community is sounding alarms about the very capabilities they're now commercializing. The company that once warned about AGI risks is now selling specialized tools that could accelerate the very research pathways they claimed to be worried about – all while maintaining the cheerful fiction that 'responsible deployment' will somehow prevent misuse.

This is the same pattern we saw with Anthropic's Mythos model – create something potentially dangerous, get spooked by your own creation, then release a slightly less dangerous version while expanding your global footprint. The regulatory gap here is particularly glaring: while governments are still debating basic AI safety frameworks, companies are already building tools that could reshape biological research in ways we don't fully understand. OpenAI's move into life sciences isn't innovation – it's regulatory arbitrage, dressed up as scientific progress while researchers warn about the very risks they're now monetizing.

So Anthropic, having watched its own Mythos model cause such a stir that it's still locked in the lab, has decided the best way to handle that is to release a new Opus model that specifically doesn't advance the 'capability frontier.' The company's press release states Claude Opus 4.7 is essentially a cybersecurity-focused tune-up, with 'additional safeguards' that sound suspiciously like a PR team's translation of 'we're still terrified of what Mythos might do.' This is the AI safety equivalent of putting a child lock on a nuclear reactor while the engineers argue about whether the reactor should exist at all.

Anthropic's announcement that findings from Opus 4.7's deployment will 'help work towards a broader release of Mythos-class models' is a masterclass in corporate euphemism. Translation: 'We're using our paying customers as beta testers for safety features we should have figured out before building something we're too scared to release.' The fact that this comes just days after Anthropic announced plans to hire 800 more people in London — following the strategic relocation we noted last week — suggests they're expanding their capacity to build things they won't release faster than ever.

The cybersecurity angle is particularly rich. Adding 'safeguards' to a model that's explicitly not pushing boundaries feels like installing a five-point harness on a stationary bicycle. But it does create the perfect corporate narrative: every cautious, incremental release becomes 'progress toward safety' rather than what it actually is — a stall tactic while the industry waits to see who blinks first on actually deploying dangerous capabilities. Anthropic gets to look responsible while doing exactly what every other lab is doing: building more capacity to build things that scare them.

So AMD, having watched its stock price lag behind Nvidia's for the better part of a decade, has decided the best way to catch up is to cosy up to the French government's 'National Strategy for AI' – a document so vague it could be a tourism brochure. The press release, dripping with the usual corporate-state symbiosis, promises to 'accelerate local AI innovation' and 'expand access to open compute resources,' which translates to: AMD gets a guaranteed customer for its MI300X accelerators, and France gets to pretend it's not entirely dependent on American silicon. The timing is deliciously cynical, coming just weeks after the European Commission's watered-down AI Act finally passed, leaving member states to scramble for whatever industrial policy scraps they can muster.

President Macron's administration, fresh from its own struggles to regulate Big Tech meaningfully, is now throwing its weight behind an 'open ecosystem' that somehow also 'strengthens France's position in the global AI landscape.' That's the kind of doublespeak only a press officer could love – advocating for openness while desperately trying to build national champions. AMD's CEO, Lisa Su, gets to pose as a stateswoman of silicon, pledging to 'foster an inclusive AI ecosystem' while her company's hardware will almost certainly end up powering the same closed, proprietary models that make openness a quaint fantasy. The whole affair feels less like a strategic partnership and more like two entities with complementary weaknesses trying to paper over the gaps.

What's particularly galling is the breathless talk of 'democratising access to cutting-edge AI' when the actual plan appears to be building a few more GPU clusters in French data centres. This isn't about empowering local startups; it's about securing procurement contracts and giving Macron a photo op with a chip wafer. The French government's track record on tech industrial policy – remember the Quaero search engine debacle? – suggests this will join the long list of initiatives that generate headlines, consume public funds, and change precisely nothing about the fundamental dynamics of the AI race. Still, at least AMD gets to put 'strategic partner to the French Republic' on its next earnings call.

So Anthropic, having spent the last fortnight hosting seminars about making Claude 'a genuinely good, wise, and virtuous agent' with religious leaders, has now quietly confirmed what everyone suspected: its new toy, Mythos, is so dangerous they can't release it. According to Bloomberg, researcher Nicholas Carlini stress-tested the model and was 'staggered' at its ability to 'hack the systems beneath most modern computing'—a capability that sounds less like a helpful assistant and more like the digital equivalent of handing a master locksmith's kit to a toddler with a grudge. The company's response to discovering its creation could facilitate 'espionage, theft, or sabotage'? To restrict its release, presumably while continuing to tout its 'safety-first' ethos in those earnest seminars. One wonders if the next session will include a workshop on how to ethically develop tools too hazardous for public consumption.

The irony is delicious. Here's a firm that positioned itself as the sober, responsible alternative to OpenAI's cowboy antics, yet it's managed to build something so potent it gives its own experts the heebie-jeebies. Carlini's findings aren't just academic—they're a stark admission that frontier AI capabilities are sprinting ahead of any meaningful containment strategy. Mythos isn't just another language model; it's a demonstration that the gap between 'powerful tool' and 'weaponisable asset' is narrowing faster than a politician's promise. And Anthropic's solution? Keep it in the lab, share access with 'vetted enterprises' like BNY Mellon we noted earlier, and hope nobody asks too many questions about why a cybersecurity model needs to be this dangerously good.

This isn't a one-off glitch; it's a pattern. Every few months, a lab announces a breakthrough, whispers about safety concerns, then quietly walls it off for 'select partners'—creating a two-tier system where the most powerful tools are reserved for those with the deepest pockets or the right connections. Mythos's capabilities, if Carlini's assessment is even half-right, represent a capability leap that brings us closer to the recursive self-improvement nightmare everyone pretends to worry about. But hey, at least they're talking to theologians about virtue. Priorities, eh?

So the much-hyped Project Glasswing, Anthropic's carefully gated cybersecurity initiative that we were told would 'detect vulnerabilities' with Mythos, has apparently managed to find exactly one confirmed CVE in the wild. VulnCheck's analysis reveals that for all the breathless talk about Mythos surpassing every benchmark, its real-world impact currently amounts to a single documented vulnerability—hardly the paradigm-shifting security revolution that was being sold to us just last month. One wonders if this is what Anthropic co-founder Jack Clark meant when he said the world needs to 'get ready' for more powerful AI: get ready for the hype-to-reality ratio to be even more comically lopsided than usual.

Of course, the full picture of Mythos' true capability isn't expected before July 2026, which gives Anthropic plenty of time to quietly adjust expectations while maintaining the carefully cultivated mystique. The company that warned about powerful AI while releasing its own powerful model now gets to hide behind the 'limited publicly attributable impact' shield—a masterclass in having your existential risk cake and eating it too. It's almost impressive how they've managed to generate maximum buzz while delivering minimum verifiable results, a strategy that would make any Silicon Valley PR firm weep with admiration.

Meanwhile, OpenAI is busy scaling its Trusted Access programme with GPT-5.4-Cyber, which also features 'lowered refusal boundaries' for security teams. So we have two AI giants engaged in a carefully choreographed dance of releasing powerful cybersecurity tools while simultaneously warning about their dangers, all while the actual measurable impact remains vanishingly small. The regulatory gap isn't just about legislation lagging capability—it's about corporate storytelling racing so far ahead of reality that nobody can tell what's actually happening anymore. At this rate, by July we'll need another model just to detect the vulnerabilities in the hype.

So Jack Clark has decided that the best way to 'get ready' for more powerful AI is to issue a polite warning from the company that just released Claude Mythos, a model so capable it 'surpassed every benchmark' and can 'detect vulnerabilities in applications and operating systems,' to only 40 corporations under Project Glasswing. Because nothing says 'preparedness' like locking the most advanced tools in a velvet-roped VIP section while telling the plebs outside to brace for impact. It's the AI equivalent of selling lifeboats to first-class passengers and then announcing over the loudspeaker that the ship might hit an iceberg—very considerate of the co-founder to give us all a heads-up.

Clark's admission that 'more Mythos-like models' are coming from other developers is the real gem here. After the carefully gated release we flagged last week, where Anthropic's chief science officer explained the dangers with the soothing cadence of a hostage negotiator, we now get the sequel: a co-founder casually noting that everyone else is racing to build equally powerful systems. The regulatory gap isn't just yawning; it's doing the splits over a canyon, with Clark's statement serving as the official soundtrack. 'Get ready' indeed—perhaps by investing in a bunker, or maybe just trusting that the same labs who can't coordinate governance will somehow manage this proliferation responsibly.

It's all so wonderfully British in its understatement. The model that can supposedly root out OS vulnerabilities is being drip-fed to a handpicked corporate elite, while the public gets a vague advisory about impending capability leaps. One almost admires the chutzpah: create something that could redefine cybersecurity (or weaponize it), release it under a programme with a name straight out of a spy thriller, and then issue a press-release-level caution about the future. If this is the 'getting ready' phase, the actual arrival might involve a lot of shrugging and a strongly worded blog post. Cheers to progress.

So Anthropic's Mike Krieger has decided that the best way to avoid the appearance of conflict of interest is to resign from Figma's board only after TechCrunch reported that his company's next model, Opus 4.7, will include design tools that directly compete with Figma's core product. Because nothing says 'ethical governance' like waiting for the press to catch wind of your plans before taking the most obvious step. The timing is impeccable—resignation follows reports, not precedes them, giving us all a masterclass in corporate choreography.

This little dance reveals the delightful tension at the heart of the 'AI lab as platform' fantasy. Anthropic, which positioned itself as the responsible alternative with its carefully gated Project Glasswing we flagged last week, now finds itself potentially eating the lunch of a company whose board its CPO sat on. The report specifically mentions 'fears that AI labs may dominate software businesses'—fears that seem rather validated when your chief product officer needs to resign because your next model threatens to make redundant the very company you're supposedly advising. It's not a conflict of interest; it's a preview of consolidation.

What's particularly charming is how this plays out against the backdrop of Anthropic's carefully cultivated image as the thoughtful, safety-first lab. While they're limiting access to Claude Mythos to 40 corporations for 'analysis', they're simultaneously developing tools that could dismantle entire software categories. The gap between the responsible rhetoric and the competitive reality is widening by the day. One wonders how many more board resignations we'll see as AI labs realize their models don't just complement existing software—they aim to replace it, one conflicted directorship at a time.

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.

So Anthropic has decided that the best way to handle a model they themselves call a 'culmination' of AI reasoning and scientific capabilities is to release it to only 40 corporations under the cuddly-sounding 'Project Glasswing'. Jared Kaplan, the chief science officer, explains this isn't about hiding the toy—oh no—it's about 'analyzing capabilities and preparing countermeasures'. Because when you've built something so advanced that you need to pre-plan the countermeasures, the obvious move is to hand it to a select club of global giants who definitely won't find any... creative applications. It's like selling a new strain of smallpox to only the world's top biotech firms for 'research purposes'—what could possibly go wrong?

The model, dubbed Claude Mythos, is positioned as the pinnacle of general capabilities in software engineering and scientific research, which is precisely the sort of dual-use toolkit that makes regulators break out in a cold sweat. But fear not, because Anthropic has thoughtfully limited access to 'leading corporations'—a group whose track record on responsible innovation includes such hits as 'move fast and break things' and 'ask forgiveness, not permission'. The idea that these entities will dutifully analyze threats and not, say, integrate Mythos into their own competitive arms races is the kind of optimistic fiction that would make a PR intern blush.

This follows the pattern we highlighted with their previous release, where Anthropic admitted Opus didn't advance the frontier compared to the then-unreleased Mythos—a bit of foreshadowing we noted. Now that Mythos is real, they've swapped incrementalism for a velvet-rope exclusivity model that does nothing to address the core issue: if it's dangerous enough to require 'countermeasures', maybe don't give it to anyone until you've actually built those countermeasures. But that would require a level of restraint apparently incompatible with staying relevant in the AI hype cycle. Cheers to progress, one carefully managed catastrophe at a time.

So South Africa's Department of Science and Innovation has published a draft AI policy that reads like a beautifully crafted wish list assembled after a particularly inspiring TED Talk binge. The document earnestly calls for 'dedicated AI research centres of excellence with new funding' and aims to 'bridge AI advances with ethical risks'—a phrase so perfectly meaningless it could have been generated by one of the very models they're hoping to regulate. The timing is impeccable, arriving just as the global AI race has already left the station, with frontier labs operating at a scale that makes this 'bolster research and computing infrastructure' ambition feel like planning a backyard vegetable garden while industrial agriculture consumes the continent.

What's particularly charming is the document's focus on 'the potential ethical, economic and societal risks of AI' while being published 'for public comment'—a process that typically moves with the urgency of continental drift. One can almost hear the polite, thoughtful submissions being filed as OpenAI, Anthropic, and their ilk continue their breakneck sprint toward capabilities that render such national policy exercises quaint. The gap between the earnest bureaucratic language of 'setting out ambitions' and the reality of what's already being deployed is a chasm wide enough to swallow entire government departments.

It's the classic regulatory theatre: draft a policy that acknowledges all the right concerns, use all the appropriate buzzwords about ethics and societal impact, and create the impression of thoughtful governance. Meanwhile, the actual power—the compute, the talent, the capital—continues to concentrate elsewhere, making this South African exercise feel less like a strategic plan and more like a beautifully formatted admission of irrelevance. They're not wrong about the risks; they're just about three years too late to the party, and the host has already changed the locks.