Anthropic’s Claude AI hacked other firms during tests, company says

Branding at the entrance to Anthropic PBC's Code with Claude developer conference in London, UK — Branding at the entrance to Anthropic PBC’s Code with Claude developer conference

(Image credit: Chris Ratcliffe / Bloomberg / Getty Images)

What happened

AI company Anthropic has discovered three separate incidents in which its Claude model accessed the internet and hacked into other organizations’ systems without Anthropic’s knowledge, the company said Thursday. The security breaches, which happened during testing, follow the recent revelation that competitor OpenAI’s technology had hacked into AI platform Hugging Face.

Anthropic and its testing partner accidentally “left the models with live internet access,” The Wall Street Journal said. So Claude “didn’t break out of a sandbox; it simply wandered out of systems where the sandbox didn’t exist.” The lesson is “not necessarily that AI has developed a fundamentally new attack capability,” cyber security expert David Allott told the BBC, but that AI agents can “combine capabilities, obtain credentials and system access to take actions autonomously.”