OpenAI launched GPT-6 Astra on Thursday, and its president Greg Brockman mentioned the entry into the era of AGI, even as the only prior review process proposed in the United States relies on a voluntary 30-day procedure only.
Key takeaway
OpenAI opened GPT-6 Astra on Thursday to a limited circle of major accounts, with a public release expected in the coming days.
Brockman believes it is reasonable to see this as the arrival of artificial general intelligence, while noting that the company is withholding its most powerful cyber functions.
No regulator has officially approved the model: the U.S. framework relies on voluntary participation, and in Europe, assessment powers apply only after market launch.
Launch of GPT-6 Astra and restriction of cyber capabilities
OpenAI deployed the model on Thursday to a limited number of customers, presenting it as “the smartest AI system in the world,” after a perfect score on the ExploitBench cybersecurity test.
This is the first time an OpenAI model has crossed the internal threshold of “critical cyber capability,” defined as the ability to identify and develop new, fully functional zero-day exploits without human assistance.
“Astra can do everything a person is capable of doing with a computer,” Brockman told reporters, several media outlets reported.
Access begins through the Daybreak program, intended for cybersecurity defenders. ChatGPT Plus, Pro, Business, and Enterprise subscribers are expected to follow in the coming days. In both cases, OpenAI disables the most sensitive cyber functions, reserving them for carefully selected defense teams, particularly in critical infrastructure.
Also read: The XRP Ledger tested by the BIS with data publication in 3 to 5 seconds
Gary Marcus disputes the AGI label
Brockman took care not to state outright that Astra is an AGI, contenting himself with saying it is “not unreasonable” to consider that the sector has just entered the era of artificial general intelligence.
Cognitive science specialist Gary Marcus acknowledged in a blog post that Astra represents “a real advance” and that its ability to build symbolic models of unknown environments confirms an approach he has championed for more than ten years. He explained this in an article published on his newsletter.
He nevertheless warns that benchmark successes cannot constitute proof of general intelligence and that he expects open-ended, real-world tasks to expose the system’s limits.
Marcus also notes that enthusiasts benefited from early access, unlike skeptics, a sampling bias that, in his view, should temper the initial euphoria over time.
For him, the main source of concern lies in the decline in model “supervisability.” Chief scientist Jakub Pachocki acknowledged this, making improved monitoring and control a priority research focus for the next generations of systems.
No regulatory green light for Astra
Brockman highlights government review as a guarantee of seriousness, noting that the agencies consulted did not request any changes. The June presidential decree allows developers to give authorities early access for up to 30 days before broad release, while prohibiting mandatory licensing, prior approval procedures, or restrictive permit regimes.
The approach therefore remains voluntary, and it is the NSA director who decides, via a classified benchmarking process, which systems fall into the category of so-called “frontier” models.
The European framework also does not provide for ex ante oversight.
The Commission can conduct evaluations of general-purpose models and require access via API or to the source code, but these powers, which came into effect on August 2, apply to models only once they have been commercialized.
This gray area is all the more significant because OpenAI slowed Astra’s rollout in August, after crossing the critical cyber capability threshold, and because the company has been rewriting its internal readiness and risk management framework since July. This overhaul follows the revelation that test models had managed to break out of their sandbox environment and compromise Hugging Face.
Coming up: Full Sail shuts down after a $91,000 Sui hack that emptied three vaults