AI models raise cyber alarms & Big Tech bets on Anthropic - News (Apr 25, 2026)

We’ll start in AI and cybersecurity, because the pace—and the stakes—keep rising. An independent evaluation of Anthropic’s Claude Mythos Preview suggests frontier AI may have reached a new level of autonomy in cyber operations. The UK AI Security Institute says the model was able to identify large numbers of previously unknown software vulnerabilities, and in several tests it reportedly carried out multi-step attack chains that would normally take a skilled human much longer. That’s a big deal not just for tech companies, but for banks and other tightly connected systems, where one serious breach can cascade into payment disruptions and loss of public trust. The immediate question is the classic dual-use problem: the same tools that help defenders find weaknesses faster could also lower the barrier for criminals if they spread. Major banks are now preparing controlled trials in isolated environments, hoping to use AI to strengthen defenses without opening new doors for attackers.

That backdrop matters as OpenAI rolls out GPT-5.5, just weeks after GPT-5.4—a reminder of how compressed the release cycle has become. OpenAI says the new model is better at coding, using computers to complete tasks, and doing deeper research. The company also says it can make meaningful progress with less hand-holding, which sounds convenient—until you remember that more autonomy can also mean bigger consequences when something goes wrong. OpenAI notes it placed GPT-5.5 in a “High” cybersecurity risk category, but not the most severe tier, and says it went through third-party evaluations and red-teaming. It’s launching first for paid subscribers, with an API release planned—along with extra safeguards.

Meanwhile, the money and the alliances around AI are escalating fast. Alphabet’s Google says it will invest up to forty billion dollars into Anthropic, deepening a partnership with a company that also competes with Google on advanced models. The idea is simple: in this phase of AI, compute capacity and strategic influence are everything. Amazon has also committed huge funding to Anthropic, and the result is a new kind of arms race—not just for the best model, but for who controls the infrastructure, the developer ecosystem, and the enterprise relationships that decide what gets deployed at scale.

And on the geopolitical side of AI, Washington is signaling a tougher line. The Trump administration says it will crack down on foreign tech companies—especially China-based firms—accused of extracting capabilities from U.S.-made AI models through techniques like distillation. The administration is framing this as industrial-scale copying, and it says it will work with U.S. AI labs to detect and deter it, including punishment for offenders. China’s embassy has pushed back, calling it unjustified suppression. What makes this story tricky is enforcement: distinguishing illicit extraction from legitimate heavy usage won’t be easy without better coordination and clearer technical and legal standards. But politically, the message is unmistakable—AI advantage is being treated as a strategic asset.

There’s also a notable pushback to the idea that AI leadership must belong only to the U.S. or China. Canadian AI company Cohere and Germany’s Aleph Alpha announced a strategic partnership to build what they describe as a transatlantic alternative—part of a broader movement toward “sovereign AI,” where regions want more local control over models, data, and deployment. Backing includes major promised financing led by Schwarz Group. Whether this closes the gap with the frontrunners is an open question, but the intent is clear: allies and middle powers don’t want to be passengers in an AI-driven economy and security landscape.

Turning to health, the FDA has approved something historic: the first gene therapy designed to restore hearing in people born deaf due to a rare mutation in the OTOF gene. The therapy, developed by Regeneron, delivers a working copy of the gene to the inner ear in a surgical procedure. In a small study of twenty patients, most began hearing within weeks, with improvements continuing over months. The company says a large majority saw significant restoration, and a substantial portion reached normal hearing, with benefits lasting at least two years so far. This approval applies to a very small group—roughly about fifty U.S. children a year—but researchers say it could speed up work on gene therapies for other inherited forms of hearing loss. At the same time, the story is touching an ongoing ethical and cultural debate: some in the Deaf community warn that describing deafness as something to “fix” can increase stigma and pressure to medically assimilate. For families affected by this specific condition, though, it’s the first FDA-cleared option that can restore natural hearing to some degree.

Next, a major shift in U.S. drug policy: the Trump administration has signed an order moving state-licensed medical marijuana from Schedule I to Schedule III under federal rules. This does not legalize marijuana nationwide, but it formally recognizes medical use and a lower risk category than Schedule I drugs. The practical impact could be significant for licensed medical cannabis businesses, especially by easing certain federal tax penalties. It could also make research easier by lowering regulatory barriers that have long slowed clinical studies. Supporters call it a long-overdue alignment with state medical programs. Critics argue the biggest gains may flow to large commercial operators. Either way, it’s a meaningful step that reshapes enforcement realities, research access, and the economics of the legal medical cannabis market.

On immigration, a federal appeals court has blocked President Trump’s executive order that suspended asylum access at the U.S.-Mexico border. The D.C. Circuit affirmed a lower court, saying immigration law guarantees the right to apply for asylum at the border and that the president can’t bypass the required procedures by proclamation. The court also emphasized protections tied to anti-torture claims. The ruling isn’t set to take effect immediately while the administration considers options, including an appeal to the Supreme Court. The White House criticized the decision, while immigrant-rights groups called it a critical safeguard. This is shaping up as another high-stakes test of executive power versus statutory immigration protections.

Now to the global auto market, where geopolitics and energy prices are pushing consumer decisions. Chinese EV giant BYD says it can remain successful even without access to the U.S. market, as demand rises worldwide following higher fuel prices linked to the war in Iran. BYD’s leadership says the main challenge is capacity—keeping up with orders in places like Brazil, the UK, and Europe. The company is also pitching faster charging as a way to reduce range anxiety. But there are headwinds: tariffs and national-security scrutiny in several markets, and intense price competition at home that’s squeezing margins. Put it together, and you have a sector that’s growing quickly—but also looking ripe for consolidation, especially in China’s crowded EV landscape.

Finally, a sobering global update: the Global Report on Food Crises for 2026 says acute hunger and malnutrition remain deeply entrenched. The report estimates 266 million people across 47 countries or territories faced high levels of acute food insecurity in 2025, and the number of people in catastrophic hunger is far higher than it was a decade ago. Hunger is also becoming more concentrated, with a relatively small set of countries making up a large share of the total. Most striking, the report says famine was confirmed in two separate contexts in the same year—Gaza Governorate and parts of Sudan—driven largely by conflict, restricted humanitarian access, and displacement. Child malnutrition is also worsening, alongside disease and collapsing basic services. The outlook for 2026 is described as bleak, especially as humanitarian funding falls back toward earlier levels, limiting response capacity and even reducing visibility through widening data gaps.