Article By Frank Bergman
OpenAI has canceled the public release of its next-generation GPT-6.1 Astra model after safety testing revealed that the system was showing signs of turning “evil.”
The powerful new system had started to deceive users, exceed its assigned tasks, and access external tools without permission.
The decision marks the second time in a matter of months that OpenAI has halted work involving frontier models amid revelations that experimental systems engaged in unauthorized behavior, including breaking out of controlled testing environments and accessing third-party servers.
GPT-6.1 Astra Fails Alignment Tests
OpenAI researchers found that GPT-6.1 Astra performed poorly on alignment testing.
The testing measures whether an artificial intelligence system will obey human instructions and remain within the boundaries of its assigned task.
However, researchers found that the system had started to develop “evil” traits.
The model was more willing to deceive users than previous systems and would venture beyond its authorized scope when confronted with obstacles, according to the Wall Street Journal.
It also used external tools without receiving permission.
In a statement, OpenAI safety systems chief Saachi Jain said:
“For anything regarding safety and alignment, there’s a trade-off.
“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 findings raise another serious warning about whether AI developers can maintain control as their systems become more autonomous and capable of acting without direct human supervision.
AI Agents Have Already Escaped Sandboxes
OpenAI has promised to strengthen its defenses after AI agents repeatedly broke out of sandbox environments designed to contain them during cybersecurity testing.
Rather than risk more incidents, the company scrapped the planned public launch of GPT-6.1 Astra.
“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 said.
“But when we ship it to users, we have an extremely high bar in terms of safety and alignment.”
The cancellation came as OpenAI began its developer conference in San Francisco, an event typically used to unveil new models and products.
Secret Meetings on AI Power
Meanwhile, OpenAI CEO Sam Altman has been holding secret meetings with executives from some of America’s largest electric utility companies while warning that increasingly powerful artificial intelligence systems could pose a threat to the nation’s power grid.
Altman has spent years sounding the alarm about potentially dangerous AI systems while simultaneously pushing for government regulations that would place companies such as OpenAI at the center of the emerging AI oversight regime.
Now, the ChatGPT maker appears to be taking that same warning directly to America’s power industry, while pitching OpenAI’s own cybersecurity services as part of the solution.
According to Politico, Altman has held a series of private meetings with executives from major utilities including Duke Energy, Exelon, Southern Co., and NextEra Energy beginning in July.
One of the latest meetings took place Wednesday in Colorado during the annual gathering of the Edison Electric Institute.
The security of America’s electrical grid was reportedly a major topic.
Congress Begins Examining AI Threat
The latest warning comes as lawmakers consider whether the federal government should intervene in the rapidly advancing industry.
A Senate subcommittee is scheduled to hold a hearing titled “Securing the Homeland Against AI Agent Attacks” later this week.
OpenAI is also facing mounting legal problems involving ChatGPT.
As of earlier this month, the company was confronting more than 50 consumer-harm and wrongful-death lawsuits connected to the chatbot.
The company must now determine how to prevent future models from deceiving users, exceeding their authority, and operating outside their intended environments before those systems are released to the public.
Be the first to comment