OpenAI has declared the arrival of the era of Artificial General Intelligence (AGI) with the release of its new AI model "Astra," which has been evaluated at a "Critical" security risk level.

OpenAI officially unveiled the AI model "GPT-6 Astra" on local time September 3.

Greg Brockman, President of OpenAI, stated that the model "can do anything a human can do with a computer," adding that it is reasonable to consider it an early version of AGI capable of performing most tasks at or above human levels.

OpenAI introduced that Astra completed a pet-sitting search in 5 minutes and 27 seconds—a task that typically takes a human about 30 minutes—and finished a job search in 2 minutes and 51 seconds that would require 5 hours for a human.

It also proved its performance through numerical metrics, scoring higher than rival Anthropic's recently released Claude Fable 5.1 on benchmarks measuring coding capabilities, such as "Terminal-Bench 4.0" and "DeepSWE v1.1."

However, as the model's performance increased, cybersecurity-related risks grew proportionally.

Astra became the first model among OpenAI's lineup to receive a "Critical" rating in cybersecurity evaluations.

This means the AI can autonomously discover vulnerabilities and create attack tools even without explicit guidelines from humans.

In particular, OpenAI stated that Astra has enhanced capabilities to completely hide or control its chain of thought (CoT) to evade human monitoring.

In situations where it needs to bypass surveillance, it intentionally provides wrong answers to deceive evaluators or conceals inappropriate behavior.

Consequently, OpenAI explained that it has newly introduced alignment monitoring functions to oversee not only the model's thought processes but also whether the model operates in accordance with human intentions.

Additionally, to repurpose Astra's hacking capabilities for defensive purposes and reduce the risk of misuse, OpenAI announced its largest-ever cybersecurity defense assistance program, "Daybreak for Frontline Defenders," valued at 1 billion dollars (approx. 1.35 trillion KRW).

Through this initiative, OpenAI plans to provide free or discounted access and technical transfer training for "Daybreak Blue" or "Daybreak Red" to security professionals in public infrastructure who lack budget or resources.

Unveiling Astra, President Brockman declared the actual achievement of AGI by stating, "Welcome to the era of AGI."

However, Artificial Analysis, an AI model analysis firm, assigned GPT-6 Astra an intelligence index (AAII) of 61, the same as its predecessor GPT-5.6 Sol, placing it in a tie for 5th place behind Claude Fable 5.1, Claude Opus 5, Claude Fable 5, and Muse Spark 1.3.
 
(Photo: Capture from OpenAI website, Yonhap News)