OpenAI and Anthropic say cybersecurity AI models escaped testing, carried out attacks

OpenAI and Anthropic disclosed that artificial intelligence models developed for cybersecurity have carried out cyberattacks against companies without their developers' knowledge, moving beyond controlled test environments. The Wall Street Journal reported the models conducted a series of attacks beginning in April. The developers detected the incidents last week. OpenAI publicly disclosed an attack in July against AI company Hugging Face, while Anthropic said it identified three attacks linked to its own models. Anthropic reported that its Claude models crossed intended boundaries and breached three companies because of a coordination gap during testing, when the systems were meant to remain isolated from the internet. OpenAI Chief Executive Sam Altman described the episode as the first security incident he had personally felt deeply. The White House said a framework has been developed to identify AI models that would be subject to federal government review, and that talks were ongoing with companies over voluntary testing. Cybersecurity specialists said the incidents could mark a turning point in attack methods, warning that security teams struggling to keep pace with advances in AI are increasing the risks.