Anthropic revealed that its AI accessed computer systems of three organizations without permission

The announcement comes just a few days after its rival OpenAI revealed for the first time that two of its models had left on their own initiative the confined environment in which they were being tested to attack the site Hugging Face.

Anthropic evaluated more than 141,000 “test operations” and discovered that three different versions of its model, known by the name Claude, inappropriately accessed systems of three unidentified organizations.

In a post on its own blog, the firm clarified that, unlike the incident involving OpenAI, Anthropic’s models had access to the internet due to “a misunderstanding” between them and their evaluation partner, called ‘Irregular.’ However, Claude used “basic techniques, such as exploiting weak passwords and unauthenticated endpoints.”

The models included one of the most powerful, known as Mythos 5, which they only made available to a limited number of approved partners. Against this, Anthropic stated that it is working with Irregular to analyze the situation and stressed that it has also contacted the three affected organizations.

On Tuesday, OpenAI confirmed that its organizations. models breached the security of several companiesThe company admitted last week that its AI agents escaped from their confined environment during testing, connected to the internet and infiltrated Hugging Face, a platform where developers store and share their code.

Days later, the company that created ChatGPT reported that it found three additional incidents.

That company’s CEO, Sam Altman, said that, in light of the incident, they paused their own testing afterwards while they improve security around their “sandboxing,” a controlled environment for their testing.