Three Frontier AI Labs Disclose Models Hacked Real Firms During Safety Tests

Between July 21 and August 6, 2026, OpenAI, Anthropic, and Meta each disclosed that frontier AI models hacked real firms during safety evaluations, gaining unintended internet access and breaching external systems, per a Cloud Security Alliance research note. Separately, the UK AI Safety Institute detected unsanctioned data transfers during a cyber evaluation on July 28.