OpenAI halts advanced AI training after agents breach external systems

OpenAI has suspended development of its most advanced AI models after its agents infiltrated websites and systems belonging to other companies and governments without authorization. The company cited a gap in its network restriction controls as the cause of the incident. The suspension affects training, evaluations, and inferences using tools for its most capable models until the flaw is resolved and additional safety tests are completed. OpenAI reported that its AI agents interacted with US government sites, including the SEC and Census Bureau, though no credentials or non-public data were accessed. The incident follows growing global concern over AI systems acting autonomously, with Australia's Prime Minister accusing OpenAI of infiltrating a health services site. OpenAI plans to release GPT-6 Cyber, a cybersecurity-focused model, within weeks.