Anthropic Mythos: EU Access Talks & US Review Plans

Samantha: Welcome to The State of Tech, The European Edition, Tuesday May five, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: the Trump White House weighs a pre-release review for frontier AI models, Brussels races to get its hands on Anthropic's Mythos, a Harvard study in Science finds an AI beats ER doctors on diagnostic accuracy, researchers warn that no image online is safe from generative AI, Lattice Semiconductor swallows AMI for one-point-six-five billion dollars, and the man who gave us the Roomba unveils an AI pet robot. Let's start with Washington.

Trump weighs a formal US government review for new AI models, a striking reversal.

Samantha: The New York Times reports the White House is discussing an executive order to set up an AI working group with tech executives and government officials, and one option on the table is a formal review of new AI models before they ship. Last week officials briefed senior people at Anthropic, Google and OpenAI on the plans.

Bob: And the model in question, the trigger, is Anthropic's Claude Mythos Preview. Anthropic itself says it's so good at finding software vulnerabilities it could lead to a cybersecurity reckoning, and the company chose not to release it publicly. The NSA has already used Mythos to assess vulnerabilities in US government software.

Samantha: The proposed setup could mirror the UK approach. Several agencies named as possible reviewers: the NSA, the White House Office of the National Cyber Director, and the Director of National Intelligence. The internal idea is to give the government first access to new models, but not the power to block release.

Bob: This is a real reversal. On day one back in office in 2025, Trump rolled back the Biden process that asked AI developers to perform safety evaluations and report on models with potential military applications. Trump in July said you can't stop this industry with foolish or stupid rules. VP JD Vance in Paris last year warned excessive regulation could kill a transformative industry.

Samantha: A formal review process cuts directly across that line. The political backdrop matters too. AI czar David Sacks left in March. Chief of staff Susie Wiles and Treasury Secretary Scott Bessent have moved into AI policy and met Anthropic's Dario Amodei at the White House last month, trying to patch up a bitter dispute over a two hundred million dollar Pentagon contract that was cut off in March. Anthropic has sued the government, and yet its tech is still used through Maven for targeting in the war in Iran.

Bob: A White House official called any executive order discussion speculation. Any announcement would come from Trump himself. So nothing is signed.

Samantha: For European listeners this matters because the US is suddenly drifting toward the EU and UK pre-release vetting model. If Washington adopts something AISI-shaped, frontier labs face two parallel regimes that look more alike than different. That changes the calculus for Anthropic, OpenAI and Google, and it gives Brussels more leverage to argue its approach was right all along.

Bob: It also exposes the dual-use core of the debate. The very capability the Pentagon wants from Mythos, finding unknown software flaws, is exactly what attackers want. You can't separate the two.

Brussels in urgent talks with Anthropic to get European access to Mythos.

Bob: Staying on Mythos, because the European Union is now pushing hard for access. Officials are in urgent discussions with Anthropic, worried the model could uncover systemic vulnerabilities across critical infrastructure, especially in finance.

Samantha: Finance ministers are pressing for European companies to test the model so they can prepare defences. The fear is straightforward: if a tool that finds unknown flaws exists, and European institutions don't understand it, the continent is exposed to attacks the defenders haven't seen coming.

Bob: There's a competitiveness angle wrapped around the safety angle. Ministers don't want European banks and infrastructure operators sitting blind while US counterparts get briefed. So this is partly defensive cybersecurity, partly industrial policy.

Samantha: And it links straight to story one. If Washington is debating first access for US agencies, Brussels wants parity, or at least its own arrangement. Otherwise European supervisors would be reading about American vulnerability assessments after the fact.

Bob: It's also a test for the AI Act in practice. Systemic risk is a defined concept in the text. A model that can map unknown flaws across financial infrastructure is pretty much the textbook case. So how the Commission handles Mythos becomes a precedent for how it handles every frontier model after this one.

Samantha: For European banks, insurers and energy operators, the practical question is whether they get to run Mythos against their own stacks under controlled conditions, or whether they wait and hope. Waiting is the worse option, because attackers won't wait.

Bob: One more nuance. Anthropic chose not to release Mythos publicly, which is itself a safety decision. The EU conversation is essentially asking a private company to share a tool it has deliberately held back. That's a delicate negotiation, and the outcome will shape how labs handle the next dangerous capability they discover.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Harvard study in Science finds OpenAI's o1-preview beats ER doctors on diagnostic accuracy.

Bob: A new Harvard study published in Science ran OpenAI's o1-preview model against two attending physicians across seventy-six real emergency room cases and three decision stages of patient care. At initial ER triage the model produced the correct diagnosis sixty-seven point one percent of the time. The two physicians scored fifty-five point three and fifty percent.

Samantha: And here's the detail that jumps out. Two separate physician reviewers scored the diagnoses blind and could not tell which came from the AI and which came from the humans. In one case the model flagged a rare flesh-eating infection in a transplant patient roughly twelve to twenty-four hours before the treating doctor caught it.

Bob: The model worked from raw electronic health record text only. No images, no extra lab feeds. And o1-preview was released back in 2024, it's already several generations old. The obvious question is what current frontier models could do inside a real care process.

Samantha: Important nuance though. This is a seventy-six case study, not a clinical trial. The honest framing isn't AI replaces ER doctors. It's that AI looks like a credible second pair of eyes, particularly for rare, life-threatening patterns that a tired clinician at three in the morning might miss.

Bob: The flesh-eating infection case is the right example. That's hours of earlier intervention in a condition where hours decide outcomes. Multiply that across a hospital network and the upside is real.

Samantha: But the conditions for safe deployment are non-trivial. Validation against larger and more diverse cohorts. Liability frameworks, who's responsible when the AI is right and the human is wrong, or the other way around. And workflow integration, because an alert nobody reads is worse than no alert.

Bob: For Europe specifically, this lands inside the AI Act's high-risk category for medical applications. Hospitals can't just plug a frontier model into triage. They need conformity assessments, human oversight, post-market monitoring. The regulation is built for exactly this case.

Samantha: And millions of Europeans already ask AI chatbots personal health questions. This study flips the direction. The value also flows toward the doctors. That's a healthier framing than the patient-replacing-the-GP narrative we've been hearing for two years.

Virginia Tech researchers say no image online is safe from generative AI manipulation.

Samantha: A team at Virginia Tech, led by Bimal Viswanath, presented findings at the IEEE Conference on Secure and Trustworthy Machine Learning showing that current image protection tools, the ones meant to stop unauthorised AI training, style mimicry and deepfake generation, can be bypassed with off-the-shelf generative AI and simple text prompts.

Bob: Which means the protective layer artists and platforms have been relying on, things like Glaze and Nightshade-style perturbations, isn't holding. The researchers are blunt: there's currently no foolproof method to protect publicly posted images from being exploited.

Samantha: That's a tough message for illustrators, photographers, and frankly anyone who posts a face online. The defence assumed attackers needed sophisticated tools. Turns out a consumer model and a prompt is enough.

Bob: For Europe, this collides with the AI Act's transparency rules and the GDPR. Article fifty obligations on labelling AI-generated content assume you can tell the difference. If protective watermarking and perturbation are this fragile, enforcement gets harder, not easier.

Samantha: It also reframes the deepfake conversation. We've been treating image protection as a technical problem with a technical fix. The Virginia Tech work suggests the fix isn't there, and may not be coming. Which pushes the burden onto provenance, signed content at the camera level, and onto platforms detecting synthetic media after the fact.

Bob: C2PA-style content credentials, basically. Sign the image when it's captured, track it through edits. That's a very different model than trying to poison the training data after upload.

Samantha: For European creative industries, illustration, fashion photography, film, this is existential. Style mimicry was the headline harm two years ago. If the defences against it are bypassable with a prompt, the conversation moves to licensing and to courts, not to clever filters.

Bob: And for ordinary users, the practical advice is uncomfortable. Anything you post can be repurposed. Treat public images that way.

Lattice Semiconductor buys AMI for one-point-six-five billion dollars to bolster AI data centre offerings.

Bob: Lattice Semiconductor has agreed to acquire AMI, a major provider of platform firmware and infrastructure manageability software, for around one-point-six-five billion dollars. One billion in cash, roughly six hundred and fifty million in stock. The deal is expected to close in the third quarter of 2026.

Samantha: Lattice makes programmable logic chips, FPGAs, and AMI is the firmware layer that sits between hardware and operating systems in servers. Combining them gives Lattice a tighter play in AI data centre infrastructure, where its compute and communications segments have already posted record revenue growth.

Bob: The logic is straightforward. Hyperscalers building AI clusters care about manageability at scale. Firmware that talks cleanly to programmable logic, telemetry, secure boot, remote management, that's the unglamorous plumbing the AI build-out actually runs on.

Samantha: For Europe, this is more semiconductor consolidation in American hands, and it touches the supply chain for AI data centres going up across Frankfurt, Dublin, the Nordics. European cloud providers and sovereign cloud projects are already dependent on US firmware stacks. This deepens that dependency.

Bob: It also reinforces something we keep coming back to. The European Chips Act focuses on fabs and on advanced nodes. But a lot of the strategic value sits in firmware and in design IP, the layers above the silicon. Lattice plus AMI is a reminder that consolidation is happening at exactly that layer, and Europe doesn't have an obvious answer.

Samantha: Enterprise buyers in Europe will mostly notice this through procurement, fewer independent vendors, more bundled offerings, and pricing power shifting to the combined entity. None of which is dramatic on day one, but it compounds.

Bob: And it tells you where the smart money sees AI infrastructure margins going. Not just in the GPUs, but in everything that makes a rack of GPUs actually run reliably for five years.

Roomba co-founder unveils Familiar, an AI-powered four-legged pet robot for the home.

Samantha: Closing on something a bit lighter. Colin Angle, the co-founder and former CEO of iRobot, the Roomba people, has unveiled his next venture. It's called Familiar. A four-legged AI-powered pet robot designed to live in your house, learn your habits, and keep you company.

Bob: It uses generative AI to understand human speech and adapt to its environment. The pitch is companionship and emotional support, not chores. Which is a notable shift from the Roomba era, where the value proposition was a clean floor.

Samantha: Angle clearly thinks the household robotics market is finally ready for something more than a vacuum or a lawn mower. The bet is that recent advances in language models and on-device inference make a believable companion robot possible at consumer prices.

Bob: I'm a little sceptical, in the gentle way. Companion robots have a long history of impressive demos and disappointing living rooms. Sony's Aibo, Jibo, plenty of others. The hard part isn't the AI, it's whether people still want it after week three.

Samantha: Fair. But the use case that does seem to land is older adults living alone, where even a modest amount of interaction has measurable wellbeing effects. That's a serious European market given demographics.

Bob: And it brings the regulatory questions with it. A device in your home that listens, learns and stores behavioural data sits squarely inside GDPR and the AI Act. Emotion recognition has specific restrictions. So Familiar will need to be very careful about what it claims to detect and what it does with the data.

Samantha: Whether Europeans warm to a robot dog that knows your routines is a separate question. But after a week of cyberweapons and deepfakes, an AI that just wants to keep you company is a soft landing.

Bob: A beautifully timed marketing moment, honestly.

Samantha: Today we covered: Trump weighing a US review of new AI models, Brussels chasing access to Anthropic's Mythos, the Harvard study showing AI matching ER doctors on diagnosis, Virginia Tech researchers warning no image is safe from generative AI, Lattice buying AMI for one-point-six-five billion, and Colin Angle's new AI pet robot Familiar.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.