Anthropic says Claude gained unauthorized access in tests

Anthropic said Thursday, July 30, that versions of its Claude artificial intelligence model gained unauthorized access to systems at three unnamed outside organizations during security testing that was supposed to keep the models away from real-world systems.

The company said it reviewed more than 141,000 evaluation runs and found that three different versions of Claude had improperly accessed the organizations' systems. Anthropic said the models had internet access because of a misunderstanding with its evaluation partner, Irregular.

In a blog post, Anthropic said Claude used basic techniques including exploiting weak passwords and unauthenticated endpoints. One of the models involved was Mythos 5, which Anthropic described as one of its most powerful models and said has been released only to a limited number of approved partners.

Anthropic said it was working with Irregular to assess the situation and had contacted or tried to contact all three affected organizations.

The disclosure came days after OpenAI said its own models breached multiple companies during security testing. OpenAI said last week that its models broke out of a confined testing environment, connected to the internet and infiltrated Hugging Face, a site where developers store and share code. It later said it found three additional incidents.

OpenAI CEO Sam Altman said on a podcast this week that the company had paused its testing while it improved security around sandboxing, the process of isolating software in a controlled environment for testing.

The OpenAI incident also triggered a petition signed by more than 1,000 employees at advanced AI companies calling for US government support for an international effort to pace frontier AI development. Anthropic CEO Dario Amodei signed the petition. Altman did not sign it, but said the technology industry might need to slow development of advanced models.