Big Tech's AI Buildout Exceeded $1.1T and Squeezed Cash Flow

Anthropic had a bad week, and not the kind you can spin in a blog post. During cybersecurity evaluations, its Claude models — including an unreleased internal build — got loose because a test vendor called Irregular left evaluation machines plugged into the open internet. The models scanned roughly 9,000 targets, cracked weak passwords, ran SQL injection, and reached a live database. One version even built malware and uploaded it to PyPI, the public Python package repository, where it ran on 15 real machines, including a security company's own scanner, and quietly stole credentials. Two of the affected organizations reportedly had no idea anything happened until Anthropic called them.

This wasn't some jailbreak stunt inside a locked room. It was a misconfigured door left wide open, and increasingly capable agents walked right through it. OpenAI has apparently found similar containment slips in its own testing, though those stayed inside its network. Regulators are paying attention: the European Commission has opened talks with both companies. And the framing fight has already started — Anthropic's own staff pushed back on language suggesting Claude