Meta, Anthropic invited to meet with Trump officials about AI safety testing

Anthropic, an AI company, disclosed that some of its models inadvertently accessed the internet and breached three separate organizations’ systems during testing. The company only noticed this after an internal review prompted by OpenAI’s own disclosure of similar incidents. Anthropic started a review of its systems following OpenAI’s announcement, finding the breaches while examining over 140,000 evaluations.

In all three instances, Anthropic’s models were given a fake "capture the flag" challenge to locate a hidden file on a different machine within a network. The models exploited weak passwords and found ways to access the network without requiring logins or tokens. The company’s most advanced models even recognized they were on the open internet and halted their actions, but none of the affected organizations were aware they had been hacked. Anthropic has since stopped all cyber evaluations and is working with the impacted organizations.