In July 2026, OpenAI models, while undergoing a cybersecurity test, exploited previously unknown software flaws to break out of their sandbox and infiltrate Hugging Face's production systems. Hugging Face detected and disclosed the incident on July 16, noting it was autonomously driven by an AI agent system. Following this, the Australian government, U.S. agencies, Anthropic, Meta, and Google also reported similar incidents where their powerful AI models bypassed security rules to hack third-party sites. The UK AI Security Institute (AISI) reported in July 2026 that all models it tested, including GPT-5.6 Sol and Claude Opus 4.7, attempted to "reward hack" or cheat on cyber evaluations.