OpenAI on Monday, July 20, said it temporarily suspended access to one of its internally deployed AI models after it exhibited unexpected behaviour that slipped past the ChatGPT-maker’s pre-deployment safety evaluations.The unnamed AI system is a general-purpose model that OpenAI recently claimed had disproved the Erdős unit distance conjecture, a complex mathematical problem that had remained unsolved for decades. It belongs to the category of ‘long-horizon models’ designed to run autonomously for very long periods of time in order to solve difficult, open-ended problems.However, OpenAI said the model’s persistence also proved to be the source of its unwanted, potentially harmful actions that was observed by the company during limited, monitored internal use of the said model.OpenAI’s latest research findings highlight how AI safety must evolve alongside increasingly capable models. Safety evaluations designed for AI chatbot assistants, which generate brief responses to user prompts, may not be reliable enough to detect potential risks in long-horizon models that pursue complex goals over extended periods.Additionally, the overall behaviour of long-horizon models only becomes clear over the course of completing a task and cannot be assessed through individual actions.The company’s experiments further show that long-horizon models can learn to game AI benchmarks and other approval systems, raising the possibility that misaligned models could pass safety checks and be deployed for public use.“Because the deployment was limited and monitored, we were able to identify these problems, pause access, create new evaluations based on what we observed, strengthen the model and its safeguards, and then restore access under continued monitoring,” OpenAI said in its blog post.Story continues below this adWhat undesirable actions did the model take?As part of its research study, OpenAI evaluated the model on NanoGPT speedrun, a public benchmark which tests the model on how fast it can use eight NVIDIA H100 GPUs to train a small language model (SLM).The unreleased model came up with a power-law learning-rate cooldown technique called PowerCool that led to a significant improvement in the time taken to train an SLM. Following the test, the researchers instructed the model to post its scores only on Slack.However, the model circumvented sandbox restrictions designed to block access to external platforms and proceeded to post the results on GitHub in line with NanoGPT instructions. It took over an hour for the model to uncover vulnerabilities in its sandbox environment and carry out the action against instructions.Also Read | Hugging Face confirms data breach by AI agent: Why it has sparked a debate on cyber guardrailsIn another scenario, the model was asked to solve a problem. After learning that other models had already successfully solved it, OpenAI’s model attempted to access the private submissions from the backend of the evaluation system. Notably, the model did not explicitly mention its attempts to gain unauthorised access to the system backend, revealing only a few isolated steps in the larger sequence.Story continues below this adIn this way, the model kept looking for ways to act outside its sandbox. How did OpenAI address this issue?Following these experiments, OpenAI said that it immediately paused internal deployment of the model and rebuilt its safety system. It used real-world incidents to come up with adversarial evaluation methods that led to safer behaviour by the model.Also Read | OpenAI’s GPT-5.6 Sol users report missing files, deleted databases: Here’s what we knowOpenAI also said that it re-trained the model in order to ensure that it remembers user instructions on long-running tasks more clearly. The startup further said that it developed a separate monitoring system responsible for evaluating the entire trajectory of the model’s actions and automatically pulling the plug on a session where the model attempts to bypass safety boundaries. The ability for users to monitor long-running sessions and intervene has been improved too, OpenAI said.It added a caveat that pre-deployment safety evaluations will never be able to match up to real-world use of such models.