AI agents run amok. Hints at sparkling new physics. How Adobe is using vibe coding. Why you should feed a cold. All that and more in this week’s edition of The Prototype. To get it in your inbox, sign up here .

That’s how long it took an AI agent to wipe out data vendor PocketOS’s entire company database–and all of its backups, according to its founder Jer Crane. That deletion had cascading effects, Crane wrote, as the company provides data services to car rental companies, impacting customer reservations, signups and other operations. (The data was eventually restored, Crane says, but not before causing a serious outage.)

When queried, the agent’s response indicated that its action violated the guardrails it was supposed to be programmed with. Crane also details the other issues that enabled this failure to happen, and it’s worth reading in full. The bottom line, Crane wrote, “This isn’t a story about one bad agent or one bad API. It’s about an entire industry building AI-agent integrations into production infrastructure faster than it’s building the safety architecture to make those integrations safe.”

This isn’t an isolated incident with runaway AI agents. There are multiple anecdotes you can find with a simple Google search. And the issues are more systemic than that–a new report from cybersecurity company Okta highlights multiple security vulnerabilities from AI agents being given access to critical systems. Though this research was focused on popular agent software OpenClaw, it highlights the danger of giving too much access to any system.

“As an AI agent gains more permissions and context, its capability increases, but so does its potential risk,” the Okta research team wrote. The report found that although sometimes safety guardrails prevailed, in other test scenarios, “agents revealed sensitive data, including secrets found in prompts or configuration files.”

A key way to rein in this behavior, the researchers concluded, is to have stricter governance controls: “As agents take on more work, they act as identities inside enterprise systems. That means they need the same kind of control plane and governance policies already in use for people and service accounts. At minimum, agent access should be limited. Long-lived tokens should be avoided. Secret storage should be centralized and secure.”

In other news, the Department of Defense has reached agreements with seven tech companies to use their AI tools to “augment warfighter decision-making in complex operational environments.”

Discovery of the Week: New Findings Might (Finally) Break The Standard Model

T he Standard Model of Physics has been one of the most successful theories in science–its predictions keep holding up despite numerous attempts by scientists to break it. This is good, in the sense that we have a useful tool for explaining how things in nature work. It’s also bad because the Standard Model is incomplete –it cannot explain some fundamental observations such as dark matter. So physicists keep trying to undermine it, hoping that new evidence could yield better frameworks for understanding the universe.

New experiments at the Large Hadron Collider, about the decay of particles called B mesons into other particles, hint that we’re close to such evidence. New data published this week strengthens suggestions from previous experiments that when a B meson decays into three other particles, the angle at which they emerge appears to be different than what the Standard Model predicts.

There’s still more work to do to confirm the results. (You may be surprised to learn that observing a type of decay that only happens to one in a billion B mesons is hard.)

But if these results hold up, there are a couple of possible reasons why, suggests Nature reporter Elizabeth Gibney: It could prove the existence of a hypothesized particle called Z’ that would be associated with a previously unknown type of physical force. Alternatively, it could be evidence of another hypothesized particle called a leptoquark.

Either way, confirming this result could pave the way for new kinds of physics, which could help future engineers and scientists build technology we can’t even think of yet.

How Vibe Coding Boosted Adobe’s Sneaks Program

O kay, so the first part of this newsletter was pretty down on AI so let me offer a bit of a counterweight: I was in Las Vegas last week to attend Adobe’s annual customer summit, where the company showed off its latest software and technology to its customers and partners. One highlight of the programming was the “ Sneaks ”–experimental projects being developed by the company’s labs. Out of many submissions, seven were presented at the conference, with the audience voting on which one they liked best.

This year was different, Adobe’s Eric Matisoff told me, thanks to “vibe coding” (using AI agents to help build software). “Last year, we had 150 ideas come in from across the company. This year because of vibe prototyping, vibe designing, vibe coding, we had over 500,” he said. Notably, he pointed out that at least one idea was submitted from every single one of Adobe’s offices, which was a first, though it also meant a lot of work for him and his team to sift through the projects.

Matisoff was clear that the projects produced using AI “are not to Adobe standards” for production–software engineers are absolutely necessary to get them up to snuff. But the rapid prototyping has been beneficial for the company, he said, pointing to his own use of such agents to code projects in “languages I never learned.” He said this has created a “flywheel effect” for the company, enabling teams to ideate and produce projects more quickly.

“We are not replacing software engineers with these technologies–we’re enabling them,” he said.

The Hot Take: We Need To Talk About Space Traffic

Each week, I ask investors for their take on tech trends within their industries. Today I’m featuring thoughts from Micah Walter-Range , who developed and maintains the S-Network Space Index, which is used to benchmark the Procure Space ETF, an exchange-traded fund focused on commercial space companies.

What tech is being overhyped right now?

Orbital Data Centers . Server farms in orbit do not make economic sense yet. However, space can benefit from the AI boom by providing data inputs via space-based sensors, enabling communication between data centers and sensors/users on Earth, and beaming power down from space to existing data centers (reducing electrical grid pressure and enabling a switch from water cooling to air cooling). Even if orbital data centers don’t materialize, space still benefits from the broader AI buildout.

What tech should more people be talking about today?

Space Traffic Management . With thousands of new satellites launching annually, the risk of collisions increases. We urgently need automated, AI-driven coordination platforms to ensure we can continue providing the services that modern life depends on (especially navigation and communication) while adding new traffic routes to commercial space stations and the Moon.

What tech are we all going to be talking about in five years?

Orbital Manufacturing . We are moving from “launching things into space” to “making things in space.” In five years, launch costs should be low enough for companies to seriously begin building the infrastructure to produce materials that are impossible to create under Earth’s gravity. Lower launch costs will also support other developments across the space industry, accelerating revenue growth for multiple companies in the Procure Space ETF.

Anthropic vs the federal government, ctd. As of today, Anthropic remains a designated “supply chain risk” as the AI giant and Department of Defense wrangle in court. Meanwhile, the Wall Street Journal reports that the White House wants to keep Anthropic from expanding access to its new model, Mythos, over concerns that the company lacks the computing power to support government use in that scenario. (Meanwhile, despite Mythos and other high-powered models demonstrating their usefulness for cybersecurity, my colleague Thomas Fox-Brewster reports that the Cybersecurity and Infrastructure Security Agency–the government’s leading digital defense office–doesn’t have access to them.)

Space-based Facebook power: Meta, Facebook’s parent company, reached an agreement to purchase a gigawatt of power from Overview Energy, which plans to build a constellation of satellites that will beam energy to solar panels on Earth at night time using near-infrared lasers. (I interviewed Overview’s CEO Marc Berte back in December .)

Water, water nowhere: Arizona has been facing dwindling supplies of water thanks to overuse and climate-driven declines. And AI projects are primed to deepen the crisis in the state. And that’s the tip of the iceberg for a water conflict that threatens to embroil all the states in the American West that rely on the Colorado river after a hot winter left the snowpacks that feed it at record lows. Many of these states are also primed to get a chunk of the $750 billion big tech companies plan to spend on infrastructure this year, further pumping up water demand.

Pro Science Tip: Yes, You Should Feed A Cold

It’s a common saying that you should “feed a cold.” And as it turns out, there’s science to back that statement up. A study published this week found that after people have eaten, the immune system’s T-cells are better able to access the nutrition they need from sugars and fats to do their job (which is to attack viruses, bacteria or cancer) than they could before they had breakfast. So if you wake up feeling like you’re catching something, it’s a good idea to not skip any meals to give your T-cells some fuel. This study didn’t assess what the participants had to eat, so next up, the researchers plan to determine if particular nutrients better prime the immune system to ward off infections.

What’s Entertaining Me This Week

A few weeks ago , I mentioned that I’d read The Tainted Cup by Robert Jackson Bennett and said I was excited to read the next book in the series. Well, now I have and A Drop of Corruption is even better than the first book. It once again features investigator Dinios Kol and his boss Ana Dolabra uncovering corruption in a biopunk fantasy setting. This time, they’re visiting the Kingdom of Yarrow, and Bennett does an excellent job of showing the rot that lays in the heart of the kings and kingdoms that the fantasy genre adores, which is contrasted with the more egalitarian–though imperfect–empire that Ana and Din serve. It’s a refreshing thing to see in a fantasy book.