Who Guards the Gate?

The US government recently blocked access to Claude Mythos because it feared the model could help hackers find and exploit serious software weaknesses.

Now it has allowed more than 100 approved US organizations to use it.

These are not ordinary companies. Many protect power systems, water networks, hospitals, banks, communication systems and government infrastructure.

The model did not suddenly become safe.

The government simply decided that keeping it away from defenders may be more dangerous than giving them controlled access.

That decision makes sense. But it also creates a much bigger question:

Who should be trusted with an AI model that can find weaknesses faster than many organizations can fix them?

Claude Mythos can study large software systems, find hidden weaknesses and help create working exploits.

Anthropic says earlier tests found serious flaws in major browsers and operating systems. Some of those weaknesses had remained hidden for many years.

Independent testing in the UK also showed that Mythos could complete a long, multi-step attack inside a simulated company network.

It still failed some difficult tests. It is not an automatic machine that can take over any system.

But the direction is clear.

Work that once required highly skilled teams, large budgets and many days may soon be done by one person using an AI model.

That is powerful for defenders.

It is equally powerful for attackers.

A normal cyberattack may steal data or lock files.

An attack on critical infrastructure can stop water pumps, delay medical care, damage transport systems or interrupt electricity and communication.