AI finds Firefox security flaws & Meta workplace tracking for agents - Tech News (Apr 22, 2026)

Let’s start with that security story. Mozilla says Anthropic’s limited-access cybersecurity model, Mythos Preview, helped it identify 271 new vulnerabilities in Firefox 150 before release. The headline isn’t just the number—it’s the signal that AI is getting seriously competitive at reading large, messy real-world codebases and spotting risky patterns. That could tilt the playing field toward defenders—if the best tools are broadly available to the people actually maintaining critical software.

Staying with AI and how it’s trained: Meta reportedly plans to collect U.S. employees’ clicks, keystrokes, mouse movements, and periodic screenshots inside certain work apps to build training data for computer-using AI agents. Meta says this isn’t for performance reviews, but the bigger issue is trust—because workplace monitoring is inherently sensitive, and the moment you cross borders, the legal and cultural constraints get much tighter.

In developer tooling, Microsoft just shipped TypeScript 7.0 Beta, and it’s a big architectural moment: the compiler has been rebuilt on a Go-based foundation. Microsoft is pitching a dramatic speed jump that could make code checks and editor features feel far snappier, especially in large projects. If this holds up in the real world, it’s one of those behind-the-scenes upgrades that quietly boosts productivity across a huge chunk of the software industry.

On the business side of AI developer tools, Anthropic got a reminder this week that pricing experiments can backfire fast. The company briefly showed changes suggesting its Claude Code tool would no longer be included at the entry Pro tier for some users, triggering immediate backlash and confusion. Anthropic says it was a limited test and reverted the page, but the takeaway is simple: in a market this competitive, uncertainty alone can push developers to shop around.

Meanwhile, OpenAI rolled out ChatGPT Images 2.0, and the standout improvement people keep pointing to is text—images that can actually include readable, correctly spelled words. That sounds small, but it’s exactly the kind of leap that makes generated images more usable for things like mockups, menus, ads, and UI concepts. The flip side is obvious: the easier it is to create convincing “real-looking” visuals, the more pressure there is on provenance, labeling, and the public’s ability to trust what they’re seeing.

Cloud infrastructure also got a practical upgrade: AWS now lets Lambda functions mount Amazon S3 Files so code can treat a bucket more like a file system instead of constantly downloading and re-uploading objects. The interesting angle isn’t the plumbing—it’s what it enables: more multi-step, stateful workflows in serverless environments, including AI pipelines where tasks need to share context without complicated workarounds.

On mobile privacy, Google is previewing Android 17 with a new Contact Picker designed to end the old all-or-nothing address book permission. The idea is to let you share just a specific contact—or even only a specific detail—rather than your entire contacts database. Android 17 is also pushing toward more “in-the-moment” location requests and clearer indicators when location is in use. It’s a privacy story, but also a fraud story: contact lists and location history are gold for profiling, scams, and data-broker ecosystems.

Now to chips, because the AI race is becoming a silicon race. A report says Google is in talks with Marvell to co-develop new AI-focused chips, potentially broadening beyond its existing design partnerships. The larger trend is Big Tech trying to reduce dependency on the most in-demand suppliers, control costs, and tailor hardware for specific workloads. It’s also supply-chain risk management—because if your product is AI, compute availability becomes strategy, not just IT.

Apple is also positioning for its next era, with Tim Cook set to step aside as CEO later this year and move into an Executive Chairman role, and hardware chief John Ternus expected to take the top job. Analysts are reading this as Apple betting its AI story will be told through devices—new hardware experiences, tight integration, and custom chips—rather than trying to outspend rivals on massive AI infrastructure. The open question is whether Apple can define a clear, differentiated AI direction fast enough to avoid becoming merely the best place to run other companies’ AI services.

Let’s pivot to energy and climate tech, where 2025 looks like a milestone year. Ember says global electricity demand growth in 2025 was fully met by clean energy, keeping fossil-fuel power essentially flat, with solar doing the heavy lifting and wind close behind. The International Energy Agency goes further, calling this an “Age of Electricity” moment as solar and battery storage scale quickly. The important nuance: demand is still expected to rise with electrification, and grids will need serious upgrades. But the direction is getting harder to ignore.

In freight, Amazon struck a deal with Einride to add dozens of heavy-duty electric trucks to its Relay network in the U.S., with Einride owning and operating them and building charging at multiple sites. That matters because long-haul trucking is one of the tougher decarbonization problems—big batteries, big routes, tight schedules. In parallel, a San Francisco startup called Humble surfaced with fresh funding and a cab-less autonomous electric freight vehicle concept. Different approaches, same message: logistics is becoming the next major battleground for electrification and automation.

Space had both a setback and a science win. Blue Origin’s New Glenn rocket has been grounded by the FAA pending a mishap investigation after a mission failed to reach its intended orbit, leaving a satellite unusable. Early flights are where rocket programs prove—or don’t prove—their reliability, and this slows an already intense launch competition. On the brighter side, the DESI collaboration finished the largest high-resolution 3D map of the universe so far, a dataset designed to pin down how cosmic expansion has changed over time—and to test whether dark energy might be evolving rather than constant.

In policy news from Europe, several EU countries are moving to bar kids under 15 from social media, and Brussels is signaling it wants a unified approach instead of a patchwork of national rules. The European Commission says an age-verification app is technically ready, using privacy-preserving proof-of-age methods. The real test will be enforcement and adoption—because the politics are loud, but the practical challenge is building something that works without turning into a surveillance tool.

And one more legal story worth tracking: newly unsealed emails in California’s antitrust case against Amazon allegedly describe pressure tactics aimed at keeping prices higher across the web. Amazon disputes the claims, but the case matters because it goes to the heart of how marketplace platforms influence pricing beyond their own storefronts—and whether “low price” branding can coexist with behind-the-scenes leverage over vendors and competitors.

Quick health and materials updates before we wrap. Researchers are reporting renewed promise for personalized mRNA-based cancer vaccines, including striking long-term survival signals in a small pancreatic cancer study—encouraging, but still early and in need of larger trials. And in a separate lab result, researchers at RMIT developed a flexible plastic film with nanoscale texture that can physically damage certain viruses on contact. If it holds up beyond the lab, it could lead to antiviral coatings for high-touch surfaces where constant chemical disinfecting is difficult.