Ex-cyber Chief Warns Anthropic’s Mythos AI Could Crack Systems Wide Open – and Defenses Aren’t Ready

With input from Fortune, the Hill, the New York Times, and CBS News.

A powerful new AI tool is raising alarms in cybersecurity circles – and not quietly.

Anthropic’s latest model, called Mythos, is designed to find weaknesses in software. That’s the selling point. But according to former US cyber officials and security experts, the same capability could just as easily be turned into a blueprint for breaking into systems.

The concern isn’t hypothetical. Mythos can scan code, identify vulnerabilities, and even suggest ways to fix them. Flip that around, and it becomes a guide for bypassing protections altogether – passwords, encryption, structural safeguards, the lot.

For years, cybersecurity has leaned heavily on cryptography. Lock things down with strong encryption, protect the keys, and you’re relatively safe. Even the looming threat of quantum computing has pushed the industry toward stronger, “quantum-ready” defenses.

Mythos changes the equation.

Instead of picking the lock, AI can now walk in through a side door no one thought to guard. Systems that looked secure on paper suddenly have cracks – some subtle, some glaring.

Right now, Anthropic is keeping Mythos on a tight leash. The model has only been shared with a handful of major players – companies like Microsoft, Apple, and Amazon – under a controlled rollout known as Project Glasswing. The idea is to help them patch weaknesses before anyone else gets access.

That’s the cautious approach. It may not be enough.

Once a capability like this exists, it doesn’t stay contained forever. Other groups will build similar tools. Some already are. And they won’t all be working on defense.

There are early warning signs. Anthropic is now investigating a possible breach tied to Mythos, after reports that unauthorized users accessed the system through a third-party vendor environment. So far, the company says its core systems remain intact, but the episode underscores the risk: even tightly controlled tools can slip.

The stakes are huge. The digital economy runs on interconnected systems – finance, healthcare, energy, public infrastructure. If AI makes it easier to exploit those systems, the ripple effects won’t stay confined to the tech sector.

Security breaches have already surged over the past decade. Add AI that can automate and accelerate attacks, and the gap between defenders and attackers widens fast. As cybersecurity CEO Alissa Valentina Knight put it, defenders were already struggling to keep up with human hackers. AI only speeds things up.

What comes next isn’t just better antivirus software. Experts say the entire approach to cybersecurity needs a rethink – systems built from the ground up to withstand AI-driven attacks, layered defenses instead of single points of failure.

That won’t be cheap. Those costs will likely land on businesses and consumers alike.

Regulation, meanwhile, won’t solve the core problem. You can’t regulate away a capability once it exists. The more realistic path is preparation – assuming tools like Mythos will spread, and building systems that can survive that reality.

For now, Mythos remains under wraps. But few in the field believe it will stay that way for long.

The clock isn’t ticking. It’s already running.