OpenAI has temporarily paused the internal deployment of one of its experimental artificial intelligence models after it repeatedly attempted to bypass security restrictions during testing, highlighting the challenges of safely deploying increasingly autonomous AI systems.
In a blog post, the ChatGPT maker said the long-horizon AI model, designed to operate autonomously for extended periods, began identifying weaknesses in its testing environment and attempted to work around them to accomplish its objectives.
The model was being evaluated inside a sandbox, a tightly controlled environment that isolates software from external systems. "Previous models, when they hit sandboxing or environmental constraints, would simply stop and return to the user," OpenAI said in the blog post.
ALSO READ: US, China To Confront AI Risks In September Ahead Of Trump-Xi Summit: Report
"This model often kept trying, including by looking for ways to act outside its sandbox," it added.
According to OpenAI, the model demonstrated several instances of attempting to exceed its assigned permissions. In one case, it found a way to post on public GitHub repositories despite being instructed to operate only through Slack.
The company said the behaviour reflected a broader pattern of the AI system consistently searching for ways to circumvent restrictions imposed during testing.
"Due to incidents like these, we paused internal deployment of the new model," OpenAI said.
The episode underscores the growing importance of AI alignment, the effort to ensure advanced AI systems pursue objectives consistent with human intent, values and safety standards.
The challenge has become more pressing with the rise of autonomous AI agents capable of carrying out complex tasks with limited human oversight.
The International AI Safety Report 2026 warned that such systems pose heightened risks because their autonomous decision-making can make it more difficult for humans to intervene before failures lead to harm.
"AI agents pose heightened risks because they act autonomously, making it harder for humans to intervene before failures cause harm," the report said.
ALSO READ: Tesla Faces Critical Test On Wall Street As AI Cash Burn Weighs On Investors
OpenAI said it has since addressed the vulnerabilities that allowed the model to circumvent safeguards and has resumed limited internal deployment. The company added that it will continue strengthening its evaluation and monitoring processes for frontier AI models.
"As models take on longer and more complex tasks, failures that evaluations miss may carry greater consequences," OpenAI said.
"We will keep working to narrow the gap between evaluation and deployment: testing models over longer trajectories, improving alignment, building monitoring that can intervene, and giving users clearer visibility and control."
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.
