Anthropic Kneecapped by Export Controls as Samsung Goes All-In on OpenAI

The week Anthropic got kneecapped by its own government, Samsung went all-in on OpenAI, and the creative industries discovered exactly how much of their work has been feeding the machine.

The story of the week

Anthropic had a rough week, and the wound was self-inflicted only in the sense that building cutting-edge AI in the United States now apparently comes with regulatory whiplash as a feature. The Trump administration abruptly ordered Anthropic to cut access to its newest models — Fable 5 and Mythos 5 — for all foreign nationals, including users inside the US and the company’s own employees. The enforcement mechanism appears to be export controls, though, as The Verge noted, “nobody understands” the rules being applied. Anthropic spent the better part of the week scrambling to restore service.

This matters beyond the Beltway. If Washington can flip a switch and block access to frontier models based on nationality — even for people physically in the country — every non-US enterprise running on American model providers just received a concrete demonstration of platform risk. The episode also prompted a broader Decoder conversation about who actually decides when AI is “too dangerous”, a question that currently has no stable institutional answer in either the US or Europe.

Separately, Ars Technica reported that AI models with advanced hacking capabilities will soon be the norm — not an edge case — making the question of access controls both more urgent and, paradoxically, harder to enforce through unilateral decree.

New models & capabilities

Samsung goes enterprise-wide with OpenAI. Samsung Electronics is deploying ChatGPT Enterprise and Codex across its global workforce, making it one of OpenAI’s largest enterprise rollouts to date. This is the same Samsung that banned internal ChatGPT use in 2023 after source code leaks. The reversal is instructive: the risk calculus has shifted from “keep it out” to “control it or lose to competitors who don’t.” OpenAI is also shipping new spend controls and usage analytics for ChatGPT Enterprise — table-stakes features that signal the product is maturing past the hype-adoption phase into the procurement-and-governance phase.

Adobe embeds AI assistants everywhere. Photoshop, Premiere, Illustrator, InDesign, and Frame.io now each have bespoke AI Assistants in public beta, while Adobe’s Firefly studio has been redesigned to offer persistent context and reusable assets across projects. This is less about a single killer feature and more about Adobe systematically closing every gap where a standalone AI tool might poach a user.

Midjourney pivots to hardware. In a genuinely unexpected move, Midjourney announced an ultrasound-based full-body scanner — its first hardware product. CEO David Holz acknowledged the distance from cat pictures to medical imaging. The company is also building a San Francisco spa, which perhaps says something about the current state of AI company ambitions.

Subquadratic attention. Miami startup Subquadratic claims to have solved the quadratic attention bottleneck that has constrained transformer architectures for nearly a decade. They are now “bringing the receipts.” If independently validated, this could materially reduce inference costs and extend context windows. Worth watching, not yet worth planning around.

Research worth knowing

OpenAI had a prolific week on the science side. Researchers used a reasoning model to identify 18 new diagnoses in previously unsolved rare childhood disease cases. A separate post detailed how GPT-5.5 Instant improves health and wellness responses through stronger reasoning and physician-informed evaluations. And in chemistry, OpenAI and Molecule.one showed a near-autonomous AI chemist using GPT-5.4 improving a key drug-making reaction.

Google’s AMIE system, meanwhile, published in Nature, matching primary care physicians in complex disease management. The healthcare convergence from multiple labs is no longer speculative.

On the safety front, Google DeepMind published its AI Control Roadmap for securing agentic systems, and OpenAI introduced Deployment Simulation, a method to predict model behaviour before release using real conversation data. Both acknowledge the same uncomfortable truth: post-deployment guardrails are not enough.

The security argument landed hard this week: a critical Copilot vulnerability allowed hackers to steal 2FA codes via a “SearchLeak” exploit, illustrating — again — that the industry’s approach to LLM security keeps failing in the same ways.

CEO watch

Barret Zoph is out at OpenAI again after just five months, having returned in January from Mira Murati’s Thinking Machines Lab. The revolving door between OpenAI, Anthropic, and their various spin-offs continues to spin at a rate that suggests either profound cultural dysfunction or a labour market so tight that loyalty is structurally impossible. Probably both.

Training data transparency: The Atlantic’s Alex Reisner built a searchable database of music used to train AI models, uncovering four datasets — two of them containing 12 million and 9 million tracks respectively. For any European company producing or licensing creative content, this is the week to check whether your catalogue appears in those datasets. The legal exposure under EU copyright law is non-trivial.

A Pew Research poll found that 49 percent of Americans now use chatbots at least occasionally — up from 33 percent in 2024 — but 63 percent think the technology is advancing too quickly. Adoption and anxiety are scaling in lockstep.

What it means for European operators

Vendor diversification is no longer theoretical. The Anthropic export-control episode is the clearest signal yet that relying on a single US-based model provider creates hard geopolitical risk. European enterprises should maintain working integrations with at least two providers and keep a clear-eyed view of what happens to their workflows if one goes dark overnight. Sovereign and open-weight models (Mistral, Llama derivatives) deserve space in the architecture even if they trail frontier performance.

Samsung’s rollout is your benchmark. If a company that banned ChatGPT three years ago is now deploying it company-wide with enterprise spend controls, the question for European mid-market firms is no longer “should we adopt?” but “do we have the governance layer to adopt safely?” OpenAI’s new usage analytics and spend controls are relevant, but so is building internal policy that survives the next model upgrade.

Creative content owners: audit now. The Atlantic’s music training-data database should prompt an immediate review of whether your IP appears in training sets. Under the EU AI Act’s transparency obligations and existing copyright directives, this is both a legal and commercial question.

The UK is moving. Google DeepMind’s partnership with the UK government on AI-accelerated housing planning is a small but telling example of public-sector AI procurement in a European-adjacent market. For firms selling into government, the playbook is becoming clearer: domain-specific, tightly scoped, politically palatable.

Tesco’s VMware exit — migrating 40,000 workloads after a 175 percent Broadcom price hike — is not AI news per se, but it is infrastructure news that matters for anyone running AI workloads on legacy virtualisation stacks. The Broadcom situation is accelerating migration timelines across European retail and logistics. If you are planning GPU-heavy infrastructure, factor in the virtualisation layer now rather than inheriting someone else’s vendor lock-in.