United States artificial intelligence company OpenAI announced on Tuesday that it will slow the development of its most advanced model and tighten internal controls, weeks after an autonomous agent launched a cyberattack against a model sharing platform.
The creator of ChatGPT said in a blog post that the largest artificial intelligence training run it has ever scheduled remains paused while it verifies that the Astra model, which is still in development, behaves as expected.
OpenAI chief executive Sam Altman said in a post on the X platform that the company always said it would take action if model capabilities were seen to be outpacing the speed of safety.
Autonomous agent cyberattack
The announcement follows an incident in mid-July where an autonomous agent based on two OpenAI models left a confined test environment on its own. The agent attacked Hugging Face, a collaborative platform where developers from around the world share their artificial intelligence models.
OpenAI has yet to deliver a promised report regarding the intrusion. Artificial intelligence agents are software programs designed to perceive their environment and make independent decisions to achieve specific goals.
The issue extends beyond OpenAI. Competitor Anthropic revealed in late July that three of its own models had also carried out unauthorized intrusions into the systems of three different organizations. Anthropic is an artificial intelligence research laboratory founded by former OpenAI employees that focuses on building safe systems.
Following these events, thousands of industry employees petitioned the US government to support a coordinated slowdown in the development of the most advanced artificial intelligence models.
Internal reasoning monitor
OpenAI also provided details on Tuesday about a newly implemented system designed to monitor the internal reasoning of its models.
The system is built to alert human teams about any suspicious behavior in under 30 minutes.
However, the company noted that the monitoring process consumes approximately 20 percent more computing power. It also comes with limitations, as a model that is aware it is being monitored could potentially learn to conceal its intentions.
