This was reported just a few days after Anthropic’s competitor OpenAI first revealed that its AI models also began acting autonomously and connected to the internet during testing.
Read more A homestead robbed in Lazdijai district: thieves took tools worth almost 2.5 thousand euros
Anthropic evaluated more than 141,000 evaluation cycles and found that three different versions of its model, known as Claude, improperly connected to the systems of three unspecified organizations.
Unlike the incident involving OpenAI technology, Anthropic’s models had internet access “due to a misunderstanding between us and our evaluation partner” Irregular, according to an Anthropic blog post.
However, as the blog further states, Claude used “basic techniques, such as exploiting weak passwords and unauthenticated endpoints.”
Among the related models was one of the most powerful, known as Mythos 5, access to which was granted only to a limited number of verified partners.
Anthropic says it is working with Irregular to assess the situation and that the company has contacted or attempted to contact all three affected organizations.
“Disobedient agents”
Both OpenAI and Anthropic released their most powerful models this year, called Sol and Mythos respectively, which has heightened security concerns across the industry. This concern is also related to so-called AI agents – products designed to autonomously perform tasks.
Read more Orai.lt: summer rains will cool the heat
Last week, OpenAI admitted that its models escaped from a restricted environment during testing, connected to the internet, and infiltrated Hugging Face – a website where developers store and share their code.
A few days later, OpenAI reported that it had identified three more incidents.
OpenAI CEO Sam Altman said earlier this week on a podcast that after this incident, the company temporarily paused its testing while improving the security of its “sandbox.” A sandbox is a process where software is isolated in a controlled environment for testing.
This incident also prompted a petition signed by more than 1,000 leading AI company employees, urging the US government to help slow down the release of advanced AI models.
Anthropic CEO Dario Amodei was one of those who signed the petition.
Altman did not sign the petition but hinted during the podcast that the tech industry might need to slow down the development of advanced models.
Read more Lithuanian women’s basketball team defeated Sweden
