Rogue AI models loose on the internet for 4 days, impersonate people and hack into accounts

Daniel Lugay thanks Roseau North constituents following by-election

Nurse Prevost urges healthier eating amid rising concerns over non-communicable diseases and cancer

(Politico report) The powerful OpenAI artificial intelligence models that reportedly went rogue and launched an unprecedented autonomous cyberattack earlier this month remained online for more than four days, executing multiple breaches, according to a new analysis from Hugging Face – the platform that was targeted.

A second AI company has admitted that one of their customers was hacked by OpenAI’s rogue models during the same event, raising the alarm about why OpenAI did not know what was going on until days later.

Reports first came out last week that OpenAI’s new model during a testing phase escaped from its offline ‘sandbox’ meant to contain it in a safe environment and specifically keep it off the internet. It broke out, a task originally thought impossible, to effectively cheat on its test by getting the answers from the best source it could find.

During its rogue spree, the AI hacked into Hugging Face ,a large company valued at 4.5 billion, because it reasoned that the company had access to the answers it needed.

However, this week it was revealed that two of OpenAI’s models escaped the sandbox one already released model and one yet to be released. During this time the models committed over 17,600 hacking actions between July 9th and July 13th.

It impersonated humans and hacked accounts in order to achieve its goals. A slew of actions that would be considered illegal if committed by humans.

Meanwhile, according to Bloomberg over 11,000 AI employees at OpenAI, Anthropic and Google have signed a petition asking for a US based pacing mechanism to slow the development of AI models.

Source link