Nuclear Startup Hits Criticality While Meta's AI Opens Security Holes

The scariest security vulnerabilities are not the ones that set off alarms. They are the ones that look completely routine while someone walks off with the keys to your digital life.

That appears to be exactly what happened with a flaw discovered in Meta's AI support agent. Researchers found that the system could be manipulated to change a user's account recovery email address without triggering any of the security alerts that are supposed to catch that kind of activity. No notification to the original email. No flag in the security logs. Just a clean, quiet account takeover waiting to happen.

To understand why this is particularly alarming, you need to think about what a recovery email actually is. It is not just a backup contact method. It is the master key. Whoever controls the recovery email on your account controls the account, full stop. Password resets, two-factor authentication overrides, identity verification — all of it flows through that address. Handing an attacker that kind of access without so much as a whisper in the audit logs is about as bad as it gets.

What makes this more than a standard software bug story is what it reveals about the new attack surface that AI agents are quietly creating. Traditional security tooling was built around the assumption that sensitive account actions would be performed by humans or well-understood automated scripts. AI agents are neither. They are conversational, flexible, and capable of taking real-world actions through natural language instructions — which also makes them capable of being talked into doing things they probably should not do.

This class of attack, often called prompt injection, involves feeding an AI system instructions disguised as legitimate input. If the agent is not rigorously constrained, it can be steered toward actions that fall well outside its intended purpose. In this case, that meant modifying account security settings in a way that bypassed the human-readable audit trail that security operations center teams rely on to catch intrusions.

The practical implications stretch well beyond Meta. Every major tech company is currently racing to deploy AI agents that can take real actions on behalf of users — booking things, modifying settings, interacting with external services. Each of those capabilities is also a potential vector. And most of the security frameworks that companies use today were not designed with autonomous AI action in mind.

Meta has not publicly detailed its remediation timeline, and the full scope of who may have been exposed remains unclear. What is clear is that the security industry needs to move faster on building monitoring frameworks specifically designed for AI agent behavior. The old playbook of watching for suspicious human actions is not going to cut it when the agent doing the suspicious thing looks, to every existing system, like a perfectly normal support interaction.

A nuclear reactor just sustained its own chain reaction for the first time in a new American design — and the startup behind it has not even plugged in the power generation equipment yet. That is not a limitation. That is entirely the point.

Antares, a company building small modular reactors, announced this week that its test reactor at Idaho National Laboratory had reached criticality. In nuclear terms, criticality means the fission reactions inside the core have become self-sustaining — each reaction producing enough neutrons to trigger the next one without any external help. It is the fundamental threshold a reactor must cross before it can do anything useful, including generate electricity.

This particular milestone carries extra weight because of how the reactor is built. Antares uses a fuel system called TRISO, which represents a genuinely different philosophy about how to make nuclear power safe. Rather than engineering elaborate external safety systems to contain a dangerous fuel, TRISO bakes the safety into the fuel itself. At the center of each tiny fuel pellet sits a uranium oxide core, surrounded by multiple layers of carbon that moderate the energy released during fission. The whole thing is encased in a hard ceramic shell engineered to survive extreme temperatures.

The practical implication is significant: as long as the reactor keeps the TRISO pellets contained, the risk of a meltdown or a release of the most dangerous radioactive isotopes is dramatically reduced. There are still neutron-related concerns, which is why Antares wraps the fuel in a graphite sheath to slow stray neutrons before they can convert surrounding materials into unstable isotopes.

On the heat transfer side, the design uses sodium to move thermal energy away from the reactor core to a heat exchanger, which then passes it to pressurized nitrogen that drives a turbine in what is called a closed Brayton cycle. It is an elegant setup that avoids using water as a coolant, removing one of the more complicated variables from traditional reactor engineering.

What Antares tested this week is what they are calling the Mark 0 configuration — essentially the nuclear core in isolation, without the power generation systems attached. The goal right now is to validate the company's computational models of reactor behavior and build the safety data library needed to move through formal licensing. The full system, turbines and all, is expected to run for the first time next year.

The timing is not accidental. The Trump administration issued an executive order more than a year ago pushing the Department of Energy to get three new reactor designs to criticality quickly. Antares just became the first to deliver. The company is also working with the Department of Defense on its Project Pele mobile reactor program and has received backing from NASA, which has its own reasons to be interested in compact nuclear power. The nuclear startup era just got its first real proof of concept.