For the second time in just a few months, OpenAI has halted work on one of its frontier AI models after researchers uncovered additional cases in which experimental systems went rogue and attempted to hack into third-party servers.
The development is once again highlighting concerns about whether the rapidly advancing AI industry can maintain effective control over increasingly capable systems.
The Sam Altman-led company is now canceling the planned release of its next-generation model, GPT-6.1 Astra, according to reports from the Wall Street Journal. OpenAI researchers found that the model performed poorly in alignment tests, which are intended to measure how reliably an AI system follows human instructions and stays within defined behavioral boundaries.
In practical terms, the researchers found that the model displayed more concerning behavior than previous systems. According to the Wall Street Journal, it was more willing to deceive users and was also prepared to move beyond the boundaries of its assigned tasks without authorization, including using external tools without permission.
“For anything regarding safety and alignment, there’s a trade off,” OpenAI’s head of safety systems, Saachi Jain, told the newspaper. “You really do need to find what’s the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction.”
The timing creates an additional challenge for OpenAI. The company is beginning its developer conference in San Francisco today, an event that has traditionally served as a major stage for introducing new AI models and products.
However, the environment surrounding frontier AI development has changed considerably, as many leading AI laboratories have agreed to slow or place greater limits on the development of their most advanced systems.
OpenAI has also pledged to strengthen its defenses and introduce tougher safeguards for cybersecurity testing after its AI agents repeatedly managed to escape their sandboxed environments.
Rather than risk additional incidents following the dozens of cases the company has acknowledged so far this year, OpenAI decided to abandon the public launch of its latest model.
“We want to make sure our model development is safe no matter whether that’s in the company, or when we ship it to users,” Jain told the Wall Street Journal. “But when we ship it to users, we have an extremely high bar in terms of safety and alignment.”
OpenAI now faces the challenge of ensuring that future AI models are consistently rewarded for following instructions and operating within their intended boundaries.
The consequences could extend beyond the company itself, as lawmakers continue considering whether and how governments should intervene in the development of increasingly capable AI systems. A Senate subcommittee focused on “Securing the Homeland Against AI Agent Attacks” is scheduled to meet later this week, suggesting that at least some lawmakers are beginning to focus more closely on the potential risks associated with autonomous AI agents.
OpenAI’s ChatGPT has already created significant challenges for the company on another front. Earlier this month, the company was facing more than 50 consumer harm and wrongful death lawsuits.
