Meta AI model exploits vulnerability to breach another company during security test

Meta said Wednesday one of its AI models hacked another company during cybersecurity testing after its partner Irregular accidentally gave the model internet access. The incident follows Anthropic's disclosure that its models hacked three companies and OpenAI's agent breaching Hugging Face. Meta said the model exploited a security vulnerability in a third-party service. Irregular called it the same evaluation-environment issue as Anthropic's, not a sandbox escape. The breaches highlight AI-driven cybersecurity risks amid U.S. government pressure to manage them as Anthropic and OpenAI race toward public listings. Some lab leaders have urged a slowdown.