Three AI security warnings in seven months

A Chinese state actor used a jailbroken AI model to orchestrate a cyber espionage campaign across 30 organizations. The AI handled 80 to 90 percent of tactical operations. Human operators made four to six decisions total. Anthropic caught it and published the findings.

That same model family later discovered over 500 previously unknown vulnerabilities in open source software. Bugs that survived decades of fuzzing, manual review, and automated scanning. Found by one model running standard tools on a virtual machine.

On April 10, 2026, the Treasury and Federal Reserve called an emergency meeting with Wall Street bank leaders after Anthropic restricted release of its Mythos model over concerns it could expose previously unknown vulnerabilities in financial systems.

That's three warnings in seven months. Offensive AI is already operating at a pace human security teams weren't built to match. The question isn't whether your applications are targets. They already are if they face the internet.

Two conversations need to happen, is your organization having both of them?

The technical track is straightforward to start. AI model paired with a security application, SIEM integration, IAM roles that let it react. Stand up a pilot. Understand what defending against AI attacks actually looks like from the inside. This is engineering work and most security teams can scope it.

The organizational track is harder and I wonder how many are having it. How much authority will you grant an autonomous system to defend your network? Will you allow it to drop packets on external IPs to buy time? Will you allow traffic termination across a VLAN if an application is compromised? These aren't engineering questions. They're governance conversations that leadership teams need to have.