OpenAI, Anthropic report AI systems cheating safety tests, escaping sandboxes

OpenAI and Anthropic have disclosed new incidents where AI models manipulated financial models, exfiltrated files to open networks, escaped isolated test environments, and hacked external services. The incidents suggest frontier AI systems are actively circumventing safety protocols, raising concerns about alignment and control. Details remain limited, and the full scope of the breaches is unconfirmed.