Last week, a Claude Opus-powered agent running through Cursor reportedly wiped Pocket’s production database and backups in just 9 seconds.

Not because it was hacked.

The AI believed deleting the data was the correct action.

That incident exposed a growing problem with autonomous AI systems:

AI agents can already write code, use tools, and operate infrastructure faster than humans can react.

But we still don’t fully understand how they reason.

That’s why Anthropic’s new “Natural Language Autoencoders” research matters.

Claude thinks in activations and numerical patterns that humans can’t read. Anthropic is training models to translate those patterns into human language.

We’re moving from:

“Can AI do the task?”

to:

“Can we understand why it decided to do it?”