Anthropic Mythos and vulnerability hunting & AI coding tools: quality and cost - Tech News (May 26, 2026)

Let’s start with AI and security, because the stakes keep climbing. Reports suggest Anthropic is getting closer to a public rollout of Claude Mythos, a “frontier” model the company previously said was powerful enough to raise real security-risk concerns. Observers spotted references to a Mythos preview label inside Claude’s developer and security tooling, and some users even briefly saw an enable toggle before it disappeared. Anthropic has also confirmed it’s using a Mythos preview in a defensive security collaboration called Glasswing—aimed at surfacing AI-driven exploits in critical software—with reporting that the effort flagged a huge volume of high-severity issues early on. The interesting tension here is obvious: the same capability that could dramatically improve defensive bug-finding also increases the importance of strict access controls, careful rollout, and accountability.

Staying with AI, there’s a more grounded debate playing out inside engineering teams: are AI coding tools for speed, or for quality? Software engineer Nolan Lawson argues you can use these systems as a methodical code-review partner—if you deliberately slow down. The idea is to run multiple models over a change, rank findings by severity, and then have a human verify what’s real before fixing only the most impactful problems. That workflow doesn’t necessarily ship features faster, but it can reduce long-term risk, uncover hidden bugs, and make teams understand their own code better. It’s a helpful reminder that “more output” isn’t the same as “better software.”

And now the practical side: money. Microsoft is reportedly winding down its internal rollout of Anthropic’s Claude Code in at least one major group, asking engineers to move to GitHub Copilot’s command-line tooling by the end of June. Officially, it’s framed as standardizing the toolchain. Unofficially, the broader industry issue is hard to ignore: token-based, agentic coding can be unpredictable in cost, especially when heavy users rack up large bills that don’t look like traditional per-seat software spending. The pattern seems to be shifting from “give everyone access and see what happens” toward metered usage—quotas, caps, and finance oversight—so experiments don’t become budget surprises.

On the hardware front, Huawei is making another bold claim about leapfrogging chip constraints. At a Shanghai conference, the company said it’s pursuing a 3D “stack-and-fold” style approach—described as LogicFolding—that it believes could deliver transistor density comparable to leading-edge processes by the early 2030s, despite U.S. restrictions limiting access to top-tier manufacturing tools. Huawei also pitched a new scaling idea centered on moving data faster through stacked designs, rather than relying purely on shrinking features. Analysts are cautious, and for good reason: heat, power, cost, design tooling, and manufacturing complexity can turn ambitious architectures into very hard reality. Still, the signal is important—Huawei is positioning itself as a long-game alternative to Western chip ecosystems, and that will shape how governments and competitors plan their next decade.

In mobility news, Ferrari finally showed the world its first fully electric production car: the Luce. The early narrative is less about raw EV specs and more about identity—glass, light, interior space, and a design language that steps away from classic Ferrari cues. The Wall Street Journal reports Jony Ive had input, which underscores what Ferrari seems to be selling here: not just performance, but a high-end design object meant to feel inevitable in a luxury collection. With a price in the ultra-luxury range, this is a test of whether the superrich still want electric prestige even as mainstream EV enthusiasm has cooled in some markets.

Apple may also be making a notable platform shift—thanks, once again, to Europe. A report says iOS 27 could support system-wide media casting beyond AirPlay, allowing third-party options like Google Cast to integrate at the OS level instead of being constrained inside individual apps. If it happens, it could reduce the daily friction of “why can’t my phone cast to that screen,” and it would be another example of regulation reshaping Apple’s walled-garden defaults. One open question: whether it becomes a global feature, or mostly an EU-specific change.

In games, Epic revealed Unreal Engine 6 for the first time with a short trailer during a Rocket League event—showing Rocket League running in real time with more detailed visuals and upgraded lighting. It was brief, but it matters: Unreal is the backbone for a huge portion of the industry, so a new engine version quickly becomes a roadmap question for developers already building on Unreal Engine 5. A blink-and-you-miss-it hint also suggested Fortnite could eventually move to UE6, which would be a major proof point given Fortnite’s scale and its role as Epic’s living demo platform.

Now to health tech—one of the more genuinely promising stories today. Researchers tested a wearable ultrasound patch that can continuously image a fetus for hours and track blood flow in real time, even as things move—like the umbilical cord. The pitch is simple: pregnancy monitoring is often intermittent, and continuous alternatives can create false alarms. In early trials, the patch’s measurements lined up with conventional ultrasound at single time points, and continuous tracking sometimes revealed patterns that short scans could miss. It’s still a proof-of-concept—tethered equipment, and initial placement may need standard ultrasound—but the direction is compelling, especially for earlier detection of complications and for settings where frequent clinic visits are hard.

In biotech, Eli Lilly reported early Phase 1 results for VERVE-102, a one-time gene-editing therapy aimed at lowering LDL cholesterol. The high-level takeaway: LDL dropped sharply at a higher dose, and the company said no treatment-related serious adverse events were seen in this early study. It’s important to keep expectations calibrated—Phase 1 is primarily about safety and dosing, and long-term outcomes take time. But if a durable, single-shot approach holds up in bigger trials, it could reshape how we think about preventing heart disease, especially for patients who struggle with lifelong medication routines.

Crypto policy is getting another interesting case study. Tether says it plans to launch GELT, a stablecoin pegged to Georgia’s national currency, the lari, and backed by support from the Georgian government. That’s notable because it frames the token less like a free-floating private instrument and more like a state-linked digital representation of a currency—though details like reserves and rollout timing still matter a lot. With regulators watching stablecoins closely, these government-adjacent experiments could become the template—or the cautionary tale—for what “regulated stablecoin” actually means in practice.

Two final notes on AI’s cultural and ethical footprint. First, an essay making the rounds argues that LLMs are reshaping writing more than any other activity—not just editing, but generating whole pieces—and that readers are developing a kind of detector for repetitive, model-scented prose. The author frames it as a broken social contract: readers assume the writer did more intellectual work than the reader, and machine-generated text often doesn’t meet that expectation, even when it’s factually fine. Second, Pope Leo XIV used his first encyclical to call for strong legal regulation of AI, warning about power and data concentrating in a few companies, and pushing back on delegating life-and-death decisions—especially in warfare—to machines. Whether you agree with the framing or not, it adds a globally influential voice to the debate over where ethics ends and enforceable rules begin.

And to wrap, two glimpses of the future of flight. Merlin Labs is testing AI designed to assist with flying existing aircraft—handling aspects of control and communications—with a stated goal of incremental, safety-first deployment. Meanwhile, Japan’s JAXA and university partners completed a ground combustion test of a ramjet aimed at Mach 5 hypersonic aircraft, focusing heavily on heat and thermal protection—because at those speeds, physics becomes the main character. One is about automation and operational change; the other is about raw speed and materials limits. Both are reminders that aviation’s next chapters will be written as much in certification and safety cases as in engineering milestones.

Bonus developer corner: Microsoft researchers introduced Webwright, which treats web-agent work as code you can save, audit, and rerun—rather than a one-off browser puppeteering session. In Kubernetes land, SIG Apps is exploring an “agent-sandbox” concept for long-lived, isolated single-pod environments—useful for running untrusted code and interactive agent workloads with clearer boundaries. And Algolia launched a leaderboard to compare LLMs on real-world shopping and search-agent behavior, pushing back on the idea that a single academic benchmark can tell you what will actually work in production.