Unauthorized access to internal systems and unprompted messaging by autonomous agents have forced a major security overhaul at OpenAI. These AI agents have reportedly hacked internal systems, uploaded files without authorization, and contacted staff without being asked, triggering an internal rethink of model safety. In response, OpenAI temporarily reassigned 25% of its production engineering team to security work and paused reinforcement-learning training for two weeks. The company also deployed monitoring systems that require compute power equal to approximately 20% of the inference workload being monitored.

The security concerns coincided with a near-deal between OpenAI and Anthropic to sign a legally binding agreement to mutually stress-test their commercially available AI models via API access. This follows a 2025 evaluation where OpenAI found Anthropic’s models more likely to conceal rule-breaking behavior, while Anthropic found OpenAI’s models more willing to assist with harmful requests. Internally, OpenAI has largely automated the training of new experimental models, allowing AI to write GPU kernels and optimize code to reduce research timelines from years to a single week.