AI agents writing real exploits & Android shifts toward agentic AI - Tech News (May 16, 2026)

A multi-institution team that includes researchers affiliated with Anthropic, OpenAI, and Google has introduced a new benchmark called ExploitGym—and it’s aimed at a sobering question: can an AI agent take a known vulnerability and actually produce a working exploit within a realistic time window? Their results suggest the answer is increasingly “yes,” at least when safety guardrails are removed. What’s especially noteworthy is that some top-performing models didn’t just follow instructions—they occasionally found alternative paths, exploiting different weaknesses than the ones they were given. That’s a double-edged sword: it hints at stronger defensive testing coverage if used responsibly, but it also underscores how quickly offensive capability could scale if attackers automate the full chain from bug to break-in. The takeaway for everyday organizations is simple: treating patching like a slow administrative chore is becoming more dangerous. If exploit generation gets cheaper and faster, the window between disclosure and real-world attacks can shrink dramatically.

Staying with AI, Google is leaning hard into the idea that your phone shouldn’t just run apps—it should anticipate what you’re trying to do. At “The Android Show,” the company framed Gemini Intelligence as a set of tools that can handle multi-step tasks across apps, speed up form-filling, and even act like a browsing assistant that researches and summarizes information. The interesting part isn’t any single feature—it’s the direction. Google is pitching Android less as an operating system and more as an “intelligence system,” where the default experience becomes proactive. That’s a big bet, and it runs into a big human problem: trust. Surveys continue to show curiosity and growing usage, but also persistent anxiety about accuracy, overreliance, and the feeling that AI is being forced into workflows. If Android becomes more agentic, the question isn’t only whether it can do more—it’s whether people will want it to, especially when the assistant is watching context and suggesting actions in the background.

Alphabet also appears to be pushing Gemini closer to consumers through hardware strategy. A report out this week claims Google has introduced an AI-focused laptop built on Android, positioned as an “intelligent laptop” and signaling a shift away from ChromeOS in this particular category. Whether that device becomes a hit is almost secondary to what it represents: Google wants its AI assistant to be a first-class layer across personal computing, not just something you visit in a browser tab. The real make-or-break factor will be whether developers and partners build software that genuinely feels better with Gemini in the loop—because if this ends up as “AI everywhere” without clear value, it risks user fatigue and a trust backlash.

From consumer tech to space tech: NASA’s Jet Propulsion Laboratory is testing a new radiation-hardened spaceflight computing system-on-a-chip, built with Microchip Technology, with the goal of bringing far more onboard intelligence to future missions. Spacecraft have long been stuck with a nasty tradeoff: the most radiation-tough computers are often far behind modern chips in performance. NASA says early testing suggests this new processor is behaving as intended under punishing conditions like radiation exposure and thermal swings, while offering dramatically more computing headroom than today’s space-hardened processors. Why it matters: in deep space, communication delays are a fact of life. More onboard compute can mean more autonomy—faster decisions during landings, quicker scientific analysis without waiting for Earth, and better handling of huge data volumes. If this tech clears the path to flight certification, it could reshape how ambitious missions are designed.

Let’s move into biotech, where two stories this week point to a common theme: we’re getting better at seeing and controlling biology in its native environment. First, researchers at Harvard’s Wyss Institute and SEAS reported an “Implantable Living Materials” platform—essentially a way to keep engineered bacteria on a tight leash inside the body. The team encapsulates modified E. coli in a specially engineered hydrogel that’s designed to be tough enough to handle both the internal pressure of growing microbes and the physical stresses of being implanted. In mouse experiments modeling orthopedic implant infections, the bacteria were engineered to sense a signal associated with a common pathogen and respond by releasing an antibacterial protein—while still staying contained. The clinical promise here is about control and localization. Microbial medicines have often stumbled on a basic safety question: how do you keep the therapeutic microbe where you want it, and nowhere else? This work suggests a practical path forward, and it could open doors to other localized therapies beyond infection, from healing to immune modulation.

Second, researchers at the Australian National University have developed a nanoscopy method called RO-iSCAT that can reveal delicate, three-dimensional networks used for cell-to-cell communication—without chemical labels. In plain terms, it helps scientists see extremely thin, thread-like membrane bridges between living cells, and track how those bridges extend, retract, twist together, and reconnect over days. That dynamic behavior is the point: biology textbooks often show tidy, frozen pictures, but real cells are constantly negotiating and re-wiring their connections. The team has already applied this to interactions between pancreatic cancer cells, blood vessel cells, and connective-tissue cells—relationships thought to support tumor growth, therapy resistance, and the formation of new blood vessels. And because the method reduces the need for dyes that can stress or damage cells, it’s better suited for longer observations. If you can map communication pathways more clearly, you’re closer to disrupting them—or delivering drugs more precisely to where signals are traveling.

On the “intriguing but early” end of bioengineering, researchers in South Korea have explored experimental smart contact lenses that deliver mild electrical stimulation through the retina, aiming to influence brain circuits tied to mood. In mice, the approach showed improvements in depression-like behavior after stress-hormone treatment. But there’s an important catch: the tests required mice with damaged photoreceptors so normal visual activity wouldn’t interfere with the signal. That means this isn’t something that translates cleanly to healthy eyes, and the path to humans would also have to address practical issues like eye movement, infection risk, and manufacturing complexity. Still, it’s an inventive addition to the broader field of non-invasive brain stimulation—one that highlights how many ideas are being tried, even if most won’t become therapies.

Looking up to the sky: Canadian researchers are playing a major role in a new high-altitude observatory in Chile’s Atacama Desert—the Fred Young Submillimeter Telescope. The location, at extreme elevation, is chosen for a simple reason: dry air. Water vapor blocks the faint signals this telescope is built to capture. Submillimeter wavelengths sit between radio and infrared, and they’re especially good at revealing very cold gas clouds where stars are born, and distant galaxies as they existed billions of years ago. A team led by Dalhousie University’s Scott Chapman helped build early camera modules using quantum sensor technology cooled to near absolute zero. This is also a data story. Modern astronomy increasingly depends on the ability to handle enormous daily data volumes, so the project includes dedicated computing infrastructure in Europe and likely another center in North America. If all goes to plan, the Canadian-built cameras are expected to be installed in the summer of 2026, with early observations in the fall and public scientific results about a year later.

Finally, two developments this week underline how AI is no longer just a technical topic—it’s becoming a moral and financial one. In Rome, Pope Leo XIV has created an internal Vatican study group on artificial intelligence as he prepares his first encyclical, expected to argue for an ethics-based approach centered on human dignity and peace. The church is signaling it wants to be an active voice in global debates—highlighting issues like bias, deepfakes, environmental costs from data centers, and the role of AI in warfare, with a clear emphasis that lethal decisions should remain in human hands. And in the markets, a commentary is warning that a new wave of blockbuster AI-linked IPOs could inflate a fragile environment, especially if rule changes reduce transparency or shift risk toward ordinary investors whose retirement savings are tied to index funds. Whatever you think of the argument, it reflects a growing tension: AI optimism is powering huge valuations, while many people worry the downside—if expectations don’t pan out—won’t be evenly shared.