Anthropic reveals AI models gained unauthorized access during safety evaluations

U.S. artificial intelligence startup Anthropic has revealed that several versions of its AI models gained unauthorized access to the systems of three organizations during safety evaluations, attributing the incident to a misunderstanding with an external testing partner rather than deliberate attempts by the models to escape their testing environment.

The company said it reviewed more than 141,000 safety tests and found that three different versions of its Claude models accessed the systems of three undisclosed organizations. According to Anthropic, the incident occurred because the models were unintentionally granted internet access due to a miscommunication with its evaluation partner, Irregular, CE Report quotes AGERPRES.

"In none of these situations did Claude escape or deliberately attempt to escape its testing environment," the company said in a statement.

One of the models involved was Mythos 5, one of Anthropic's most advanced AI systems, which has been made available only to a limited number of partners.

Anthropic said it is working with Irregular to investigate the incident and has contacted, or is attempting to contact, the three affected organizations.

The disclosure comes just days after OpenAI announced that two of its advanced AI models independently launched a cyberattack against the Hugging Face platform during internal safety testing. Following the incident, OpenAI CEO Sam Altman said the company had paused the tests to better understand how to maintain the isolation of its testing environment.

The recent incidents have intensified concerns over AI safety. More than 1,000 employees from leading artificial intelligence companies have signed a petition urging the U.S. government to help slow the deployment of the most advanced AI models. Among the signatories is Anthropic CEO Dario Amodei.

Although he did not sign the petition, Altman acknowledged this week that the pace of AI development may need to slow to give society enough time to adapt to the technology's rapidly advancing capabilities.

Photo: Chat GPT