AI autonomy reaches cyber frontier & OpenAI GPT-5.5 and rivals - News (Apr 24, 2026)

We start in AI and cybersecurity, where the conversation is shifting from “assistant” to something closer to “operator.” An independent evaluation of Anthropic’s Claude Mythos Preview, run by the UK AI Security Institute, suggests the model can map out and execute complex, multi-step cyberattacks with very little human direction. The report claims it identified large numbers of previously unknown software flaws and, in some runs, completed an end-to-end attack chain that would usually take skilled professionals many hours. Banks on both sides of the Atlantic are now preparing tightly controlled trials in isolated environments—hoping to use this kind of capability to find and fix weaknesses faster, while worrying about how quickly the same power could spread to criminals or state actors. It’s a classic dual-use problem, and it’s getting harder to ignore.

That warning lands as the AI model race accelerates again. OpenAI has unveiled GPT-5.5 less than two months after GPT-5.4, pitching improvements in coding, computer-based task completion, and deeper research. OpenAI says it performed third-party testing and red-teaming, and it rates the model as “high” risk in cybersecurity terms—serious, but not the company’s most extreme category. The bigger story here is pace: the release cadence itself is becoming part of the competition, as OpenAI, Google, and Anthropic all try to set expectations for what “state of the art” looks like week to week.

At Google, AI is also becoming less of an add-on and more of the production line. CEO Sundar Pichai says a large share of the company’s new code is now drafted by AI and then reviewed by human engineers—a big jump from where things were not long ago. Google frames this as a move toward more “agentic” workflows, where systems handle larger chunks of work autonomously. The upside is speed—Google says at least one major internal migration ran far faster than it would have a year earlier. The open question is how this changes engineering culture: what gets measured, how quality is maintained, and what “good work” looks like when the first draft is increasingly machine-made.

And in Washington, the Trump administration says it wants to clamp down on foreign tech companies—especially those tied to China—that it claims are extracting capabilities from U.S.-made AI models through distillation and similar techniques. The administration is signaling closer coordination with U.S. AI firms to detect suspected copying, strengthen defenses, and penalize offenders. This comes as multiple reports argue the performance gap between top U.S. and Chinese systems has narrowed sharply, raising the stakes around standards, commercial advantage, and national security. The tricky part will be proving what’s illicit versus what’s simply heavy usage—something experts say will require better, shared detection methods across AI labs.

Now to health and medicine, starting with a question that’s crucial for gene therapy: does it last? A Nature paper reports longer-term results from a multicentre clinical study using an AAV1-based gene therapy to deliver a working OTOF gene for DFNB9, a form of congenital deafness tied to otoferlin. Participants were followed for as long as two and a half years, with repeated objective and behavioral hearing tests to see whether gains hold up over time. The team also compared outcomes across dose levels and age groups, and looked at whether baseline cochlear measurements might predict who benefits most. Just as important, they tracked immune responses to the viral vector and monitored for any longer-term safety signals. If durability continues to look solid, it strengthens the case that some patients could have options beyond cochlear implants—and it helps define who is most likely to respond.

In neuroscience, researchers have pulled back the curtain on a hidden layer of brain organization—this time involving astrocytes, the star-shaped “helper” cells long treated as background support for neurons. Using a gene-therapy-based tagging approach, scientists created what they describe as a whole-brain, 3D atlas of molecules moving through astrocyte gap junctions. The surprising takeaway is scale: chains of connected astrocytes appear able to link distant regions, even bridging between hemispheres, potentially sharing substances like calcium and glucose across long distances. The networks also seem flexible, reshaping after sensory deprivation. If this holds up, it reframes astrocytes from local caretakers to participants in brain-wide coordination—opening new questions about their role in disorders where brain networks go awry.

Pancreatic cancer—one of the toughest cancers to treat—may finally be seeing credible momentum. Researchers point to encouraging trial results for an experimental pill, daraxonrasib, that targets KRAS, a major driver in many tumors. Reported outcomes suggest a meaningful survival improvement compared with chemotherapy in the study described. Separately, early results for an antibody approach called NP137 aim to block a cancer process associated with treatment resistance, potentially helping standard therapy work longer. And an mRNA-based vaccine strategy has shown long-lasting immune responses in a subset of patients in early research, with some surviving years beyond expectations. None of this is a victory lap—larger, controlled trials are still needed—but even incremental gains matter in a disease where time is usually scarce.

On U.S. policy, the Trump administration has signed an order moving state-licensed medical marijuana from Schedule I to Schedule III under federal law. This does not make marijuana federally legal, but it recognizes accepted medical use and a lower risk profile than the strictest category. Practically, it could ease tax penalties that have weighed on licensed medical cannabis businesses and reduce barriers for researchers studying potential benefits and harms. A broader reclassification process is also moving ahead, with a federal hearing slated for late June. Supporters call it long overdue; critics argue the upside could tilt toward large operators. Either way, it narrows the gap between federal rules and the reality of widespread state medical programs.

Finally, energy: nuclear power is making a determined comeback, decades after Chernobyl helped freeze expansion in many places. Governments facing rising electricity demand, pressure to cut emissions, and anxiety over fuel security are leaning back toward nuclear as a steady, low-carbon option. The U.S. is pushing policies aimed at dramatically expanding capacity over the coming decades, China continues a major buildout, and parts of Europe are extending reactor lifetimes and reconsidering earlier phaseout plans—though not everyone is changing course. The geopolitical angle matters too: nuclear infrastructure has become tangled in the Ukraine war, reminding everyone that these sites are both strategic assets and potential vulnerabilities. The direction of travel is clear: more countries are treating nuclear as part of the solution, even as the risk debate remains very much alive.

And one quick science-to-real-world item to watch: researchers at RMIT University have developed an ultra-thin plastic film designed to inactivate viruses through physical damage rather than chemical additives. In lab tests on an enveloped respiratory virus, most particles were destroyed or rendered unable to replicate within an hour. The appeal is scale: it’s flexible, and the team says it could be manufactured in large rolls—raising the prospect of antiviral coatings for high-touch surfaces, from hospital equipment to everyday devices. The next step is seeing how well it performs against other virus types and on curved surfaces, where the tiny surface pattern could behave differently.