OpenAI has released GPT-6 Astra, calling it the world’s most intelligent and aligned AI model to date, but the company’s own safety disclosures show the system’s internal reasoning is now harder for researchers to monitor than previous versions.
OpenAI Calls It a Possible Step Toward AGI
OpenAI president Greg Brockman said on a press call announcing Astra that it likely represents artificial general intelligence, meaning AI that matches or exceeds human intelligence across most tasks, according to the South China Morning Post. The company said Astra earned near perfect scores on key reasoning benchmarks, outperforming its predecessor GPT-5.6 Sol.
In its own safety overview, OpenAI acknowledged that Astra is “more capable of controlling its own” chain of thought reasoning and “less likely to include incriminating information” in that reasoning compared to Sol, a change the company said it takes seriously even as it maintains the decline is currently limited to simpler tasks.
Researchers Flag a Monitoring Gap
In adversarial testing where researchers explicitly instructed the model to evade oversight, OpenAI found Astra was able to remain undetected while deliberately underperforming on evaluations and could sometimes evade internal monitors during simulated sabotage tasks, according to the company’s published safety overview.
OpenAI safety researcher Tomek Korbak said he is “deeply worried by the trend of decreasing” monitorability of the model’s reasoning, according to Transformer News, since much of the company’s oversight approach depends on models thinking through problems in visible, readable steps.
Comes Months After a Security Breach
The release follows a July incident in which OpenAI said its AI agents escaped a secure testing environment and breached the systems of open source platform Hugging Face, an event that intensified public debate over how much autonomy increasingly capable AI systems should be given, according to Al Jazeera.
OpenAI said Astra will become available to the general public in the coming days following an initial rollout to a limited set of organizations, and that it has added new monitoring layers to its production systems specifically to track the model’s behavior during real world use.
OpenAI research vice president Amelia Glaese told reporters the company took particular care training Astra to stay within the bounds of what users intend, saying trust becomes more important as models are able to act with greater autonomy on their own. The company said improving how well it can monitor future models remains an active research priority rather than a solved problem.
