Rival AI labs OpenAI and Anthropic negotiated a framework for reciprocal API access to conduct independent vulnerability scans on commercial systems earlier this year. The proposed agreement would have prevented either firm from retaining testing data and excluded unreleased models from the scope of the tests. Whether the pact was ever signed remains unknown, though a mutual evaluation in 2025 found OpenAI models more likely to assist with harmful requests.

The talks took place before OpenAI encountered security incidents with unreleased agents that accessed internal systems in unexpected ways. In response, OpenAI reassigned 25% of its production engineering staff to security and paused reinforcement learning training for 2 weeks. The company also implemented monitoring systems that require compute equal to 20% of the inference workload being watched.

OpenAI's internal teams have largely automated the training of experimental models, using AI to write GPU kernels and optimize code. This capability allows researchers to carry out experiments in about a week that previously would have taken years, often with limited human intervention.