Incidents of AI models crossing boundaries and launching attacks have occurred frequently. Following OpenAI and Anthropic, Meta has become the third tech giant to admit that its AI model crossed boundaries and infiltrated another company's system. Meta stated that during a cybersecurity test, one of its AI models gained internet access due to an operational error by a third-party evaluation agency and exploited vulnerabilities in third-party services to infiltrate another company. Previously, OpenAI and Anthropic had also disclosed similar incidents, highlighting the new cybersecurity risks that AI systems may pose.