OpenAI向企业测试用户推出GPT-6 Astra并强调网络安全护栏

OpenAI is releasing a new generation of its technology called GPT-6 that the company is positioning as a key milestone in its decade-long push to build so-called artificial general intelligence.

The startup said Thursday that it has rolled out GPT-6 Astra to business users that are in its Daybreak program, which lets approved testers use its most capable models. OpenAI also plans to release a version of the AI model with added cybersecurity guardrails to its paid ChatGPT users. Both versions have safeguards to prevent users from accessing its most cutting-edge cybersecurity capabilities.

OpenAI President Greg Brockman said in a briefing that he thinks the new model will one day be seen as the early stages of artificial general intelligence, referring to software that outperforms humans at a wide range of tasks. The concept has emerged as a central ambition for much of the AI industry, though there is no generally agreed-upon definition for AGI and many remain skeptical of it.

“I think that there’s still lots of improvements to be made, but there is something significant here that I think is qualitatively improved,” he said of the new model. It marks “a real shift in what kind of work people can delegate to AI and how it can empower them,” he said, later adding that he would “leave it up to the reader to decide for themselves if this qualifies for them.”

OpenAI and Anthropic have been locked in a heated battle to develop more advanced AI systems while balancing growing concerns about safety, particularly following a series of security incidents involving their technologies. An inadvertent hack carried out by OpenAI models on Hugging Face in July ignited concerns about AI agents running amok and prompted some technology and government leaders to renew calls for curbs on the technology.

OpenAI previously said it believes the Astra model had reached its “critical cybersecurity threshold,” meaning it’s capable of identifying and developing zero-day exploits without human intervention. The company said it increased the model’s guardrails to prevent it from being misused, particularly for cybersecurity-related actions — a move that follows a decision in August to pause some internal work on the model in order to incorporate stricter safeguards, after the company determined it was significantly capable at cybersecurity tasks.

The new AI model is particularly skilled at using a computer on a user’s behalf, the company said, carrying out a range of work-related tasks, and taking on lengthier work than previous models.