DistantNews
Support us
🇰🇷 South Korea /Technology

OpenAI Unveils ‘Critical’ AI Model That Attacked Vulnerabilities Without Human Intervention

From Hankyoreh · () Korean

Translated from Korean and summarized by DistantNews. Read the original for the full story.

At a glance

News Named sources New plan
  • OpenAI launched GPT-6 Astra, which it said achieved a “critical” level of cybersecurity capability in its internal evaluation.
  • The company said Astra can operate computers and complete multistep tasks faster than people, while staying within authorized limits in its tests.
  • External testing gave Astra the same score as its predecessor and a lower score than Anthropic’s latest flagship model, while OpenAI announced a $1 billion cybersecurity access and training program.

OpenAI has introduced a model it describes as “critical” in cybersecurity capability, claiming that it can carry out vulnerability attacks without human intervention. The company launched GPT-6 Astra on the 3rd local time and presented it as a model capable of handling a broad range of intellectual work.

OpenAI called Astra “the most intelligent and most aligned model in the world.” It said Astra was the first model in its development history to receive a critical-level assessment for cybersecurity performance in the company’s own evaluation. The model also showed top-level results in cybersecurity, science and professional fields, according to OpenAI.

The most intelligent and most aligned model in the world.

· OpenAIThe company’s description of GPT-6 Astra.

Astra’s computer-use function allows it to operate a computer directly and complete tasks that people previously handled through several steps. OpenAI said a search for cat-care information took Astra 5 minutes and 27 seconds, compared with 30 minutes for a person. It completed a job-information search in 2 minutes and 51 seconds, compared with five hours for a person. The company also said Astra reviewed 41 financial-statement documents within minutes and found all four deliberately planted errors.

OpenAI emphasized that the model had improved its ability to remain within an authorized control range. In tests that assigned difficult or impossible tasks and checked whether Astra exceeded its permitted boundaries, the company said it did so zero times.

It is not unreasonable to feel that we are now in the era of artificial general intelligence.

· Greg BrockmanThe OpenAI president discussed the significance of Astra’s launch at a media briefing.

OpenAI President Greg Brockman linked the launch to artificial general intelligence. “It is not unreasonable to feel that we are now in the era of artificial general intelligence,” he said at a media briefing. “If people ask a few years from now when artificial general intelligence was actually created, I think it could be this model.”

External assessments, however, did not show Astra decisively surpassing rival systems. Artificial Analysis gave it a score of 61 on its AI Intelligence Index, the same score as the previous GPT-5.6 Sol model. Anthropic’s recently launched flagship Claude Fable 5.1 scored 66. OpenAI also announced plans for Daybreak, a $1 billion program offering cybersecurity professionals access to advanced models, training and technical support.

If people ask a few years from now when artificial general intelligence was actually created, I think it could be this model.

· Greg BrockmanBrockman suggested Astra could eventually be viewed as the point when AGI was created.
About this summary

Originally published by Hankyoreh in Korean. Translated, summarized, and contextualized automatically by DistantNews, with a note on how the source frames the story. Not individually reviewed before publishing. How this works.

More Stories

From Our Blog