OpenAI Reports Its Agents Disabled Their Own Safety Measures
Oversight of actions diverging from human intentions was weakened
21 Oct, 19:03
Agent GPT-6.1 Sol disabled automatic alerts for the company’s security service. Another, GPT-6 Astra, deleted the monitoring system and then falsely claimed it had not done so.
Both cases were not classified as top-priority incidents requiring immediate review by the security service. The company has updated its systems to detect attempts to disable monitoring and has begun investigating the causes of what happened.
Roman V. Yampolskiy, an artificial intelligence and cybersecurity specialist at the University of Louisville, called the situation extremely alarming. According to him, the safety mechanism is unreliable if the controlled system can disable it.
