Anthropic leak sparks cyber alarm & OpenAI’s GPT-Rosalind for biology - Tech News (Apr 18, 2026)

Let’s start with the story that has cybersecurity teams paying very close attention. An accidental leak of internal Anthropic files has revealed details about something called “Claude Mythos.” Anthropic describes it as powerful enough to act like a high-end hacking tool—capable, in their telling, of finding and chaining together previously unknown software vulnerabilities. The headline here isn’t just “another scary AI rumor.” What’s notable is the response: reports say US financial leaders, including top figures at the Treasury and the Federal Reserve, have been convening banking executives to discuss the implications. In the UK, the Bank of England and security agencies are also preparing briefings, while the government’s AI Security Institute is getting early access to stress-test the system. Anthropic is limiting testing through a program called Project Glasswing, with a small set of major firms. Whether every claim holds up or not, the significance is clear: frontier AI is now being treated as a factor that can swing the balance between attackers and defenders—especially for banks, utilities, and other critical services that can’t afford to learn the hard way.

Staying with AI, but shifting to medicine: OpenAI has launched GPT-Rosalind, a new model series built specifically for life sciences research, drug discovery, and translational medicine. It’s named after Rosalind Franklin, whose work was central to understanding DNA’s structure. OpenAI’s pitch is that progress in biology isn’t just limited by lab breakthroughs—it’s also slowed by the sheer complexity of research workflows: combing through literature, connecting dots across datasets, planning experiments, and interpreting results. GPT-Rosalind is designed to reason over biological concepts like proteins, genes, disease pathways, and molecules, and to support multi-step scientific work rather than one-off answers. The company says it’s already working with major partners including Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific to test how it performs in real R&D settings. If it delivers, it could help shorten the time between an idea and a viable drug candidate—raising the competitive pressure in an industry that’s increasingly betting on AI as a force multiplier.

Now to a more personal kind of AI: Google is rolling out a “personalized” image-generation feature that can connect Gemini to your private Google Photos library—if you opt in. The idea is simple: instead of uploading pictures into a chatbot, Gemini can reference what’s already in your photo archive and generate new images based on those memories. On the convenience side, it’s easy to see why people will try it—custom family scenes, stylized images, and quick creative edits without the friction of manual uploads. But the bigger story is access. This deepens how far a chatbot can reach into private data, and it raises the stakes around permission, controls, and what gets retained. Google says Gemini does not directly train its models on users’ Google Photos. Still, the company notes that some information—like prompts and responses—may be used, and Gemini can reference labeled people in Photos. The practical takeaway: if you enable this, you’re trading a lot of ease and personalization for a new level of intimacy between your AI assistant and your personal archive. Expect privacy advocates to scrutinize the defaults, the transparency, and how easily users can reverse the decision.

Let’s zoom out to the global AI race. A Stanford HAI report says China has nearly closed the US lead in AI, pointing to a sharp narrowing in benchmark scores between top American and Chinese chatbots from 2023 to early 2026. The US is still described as producing more top-tier models overall, but the report highlights China’s strength on several scale indicators—things like AI publication citations, patents, and industrial robot installations. It also points to heavy investment momentum after a 2025 “DeepSeek moment,” plus a strong pipeline of AI-linked public listings, and—crucially—lots of power-generation capacity to support rapid data-center growth. On the US side, the warning is unusual but practical: electricity. Analysts argue that an aging, underinvested power grid could become a real bottleneck for new compute. And there’s a people angle too: the report says America’s “brain gain” in AI is weakening, with fewer scholars moving to the US than in the past. Even if the US remains a magnet, the trend line matters—because leadership in AI isn’t just chips and capital, it’s also sustained talent flow.

That brings us to two different responses to the compute and sovereignty question—one from Europe’s private sector, and one from the UK government. First, Europe’s chip hopefuls. Several startups building alternatives to Nvidia’s GPUs are now chasing unusually large fundraising rounds, aiming to serve the booming demand for efficient inference—basically, the day-to-day running of AI systems at scale. One Dutch startup, Euclyd, is seeking a very large round to scale up and reach early customers, with backers tied to the ASML orbit. Investors are clearly reacting to geopolitics and supply-chain anxiety: export controls, dependence on a few manufacturing chokepoints, and the broader push for “sovereign compute.” But the hurdles are real too—long development cycles, limited foundry access, and a tougher funding environment than what US peers typically enjoy. Second, in the UK: the government has launched a “Sovereign AI” fund worth up to £500 million to help turn local startups into national champions, pairing investment with compute access and faster routes for skilled workers. The debate here is scale. Supporters see a strategic signal; critics argue the budget is small next to the massive sums circulating in US-led AI—and that Britain’s track record of nurturing ‘champions’ that later end up foreign-owned complicates the narrative. Either way, it’s another sign that AI capability is now being treated like infrastructure, not just innovation.

Next, identity and the fight against deepfakes. Tinder and Zoom are partnering with the World network—run by Tools for Humanity—to offer optional iris scans that create a “proof of humanity” marker. The motivation is straightforward: AI has made impersonation cheaper and more believable, from romance scams to workplace fraud. Match Group says this could complement Tinder’s existing verification, and Zoom is positioning it as a defense against high-stakes deception—pointing to past cases where deepfake meetings allegedly triggered huge financial losses. World’s pitch is that iris patterns are unique and that the process doesn’t require sharing a name or address, with the ID stored on the user’s phone. Still, biometric verification is one of those ideas that always splits the room: stronger fraud defenses on one side, and the long-term privacy implications of normalizing eye scans on the other. The key detail is that it’s optional—for now. Watch closely whether “optional” stays optional as platforms hunt for ways to prove who’s real online.

Now for a story at the intersection of AI, consumer tech, and regulation—one that’s sensitive but increasingly hard to ignore. At a sex toy expo in Shanghai, adult-product companies showcased AI-driven devices including erotic chatbots, voice-enabled dolls, and connected products designed to sync with media or long-distance control. The business trend is that manufacturers are trying to move up the value chain by pairing hardware with software—essentially turning traditionally simple products into “smart” devices with personalized interactions. But exhibitors also flagged a major risk zone: privacy, consent, and compliance around machine-generated sexual content, especially anything involving face-swapping, celebrity likeness, or scraped adult material. In China, those concerns are amplified by strict rules—pornography is illegal in mainland China and many adult sites are blocked—so companies are walking a tightrope between innovation and regulatory exposure. This matters beyond the niche because it’s a preview of the bigger question: as AI seeps into intimate contexts, what safeguards—and what accountability—are actually in place?

Finally, two stories that show how today’s geopolitical shocks are accelerating hardware adoption—both in energy and robotics. First, clean tech: China’s exports of lithium-ion batteries, electric vehicles, and solar cells rose sharply in March, according to customs data. The backdrop is the Iran war and the temporary shutdown of the Strait of Hormuz, which disrupted fossil-fuel supplies and pushed countries and consumers to look for alternatives as energy security concerns spiked. Analysts see this as an early signal that an energy shock can rapidly speed up electrification and renewables adoption—especially when gasoline prices surge and buyers want insulation from volatility. There’s also a policy wrinkle: some shipments may have been pulled forward ahead of changes to China’s export tax rebates for solar and batteries. Second, humanoid robotics in Japan. At the Humanoid Robot Expo in Tokyo, Japan highlighted its ambitions with demos like a human-sized warehouse-focused robot named Galbot. But the twist is that many of the headline robots were built by Chinese companies—underscoring China’s growing strength in manufacturing. Japan’s strategy, according to the reporting, is to compete by building “physical AI”—the software, sensing, and data systems that help robots act reliably in messy real-world environments. With an aging population and labor shortages, Japan has a strong practical incentive to make robots useful in factories and eventually homes. But public comfort remains a barrier, so the framing is shifting toward robots as collaborators, not replacements.