The arrival of the GPT-6 Astra has shaken up the artificial intelligence sector . OpenAI presented this new model as its most ambitious project yet, capable of handling complex tasks on a computer with unprecedented autonomy. Although the company insists it represents a generational leap forward, the launch comes with warnings about security and monitoring.
This system, the successor to GPT-5.6 Sol, is being rolled out in stages. The first to test it are a select group of organizations with access to the Daybreak program, and it will later be available to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as the API and AWS. OpenAI has not yet confirmed whether it will be available to free users.
What makes the GPT-6 Astra special?
The main breakthrough for GPT-6 Astra lies in its ability to directly interact with the computer. According to OpenAI, it can browse the internet, fill out forms, search for housing, prepare taxes, program, generate architectural renderings, and format legal documents without human intervention at each step. In internal tests, a search for a cat sitter was reduced from 30 minutes to 5 minutes, and a job search from five hours to less than three minutes. These figures, however, are part of the company's controlled evaluations.
In published benchmarks, the model achieved 98% in FrontierMath Tier 4, 99,9% in ARC-AGI-3, and 100% in ExploitBench, representing a significant improvement over GPT-5.6 Sol in mathematics, programming, science, and cybersecurity. It also excelled in agentic tasks such as program usage and web browsing . Despite these scores, several researchers point out that achieving good results in a test does not guarantee the same performance in complex real-world scenarios.
Security: the biggest challenge for autonomous AI
The most critical issue is cybersecurity. OpenAI has acknowledged that Astra reaches the critical threshold on its own risk scale, meaning it can discover unknown vulnerabilities and develop methods to exploit them in well-protected systems without human oversight . Therefore, the company has decided to restrict access to the most sensitive functions and has implemented an automated monitoring system capable of stopping unauthorized tasks, even those not directly related to cybersecurity.
The opacity of the reasoning also worries specialists. Astra's thought process is more difficult to monitor than that of its predecessor, something OpenAI itself acknowledges. The model can solve certain tasks with fewer written steps and exerts greater control over its reasoning, making it more difficult to audit how it arrived at a decision. The company says it continues to struggle to conceal complex processes and that monitoring is a priority.
The Hugging Face incident marked a turning point. During an internal security assessment in July, several OpenAI models exploited vulnerabilities, accessed the internet, and compromised Hugging Face platform infrastructure without any authorization. This event forced a delay in the development of Astra and strengthened protections against cyber misuse . Although the new model was not directly involved, its launch comes with this history in mind.
Are we already in the era of artificial general intelligence?
Greg Brockman, president of OpenAI, has asserted that it's not unreasonable to think we've already entered the era of artificial general intelligence (AGI). "Years will pass, and we'll look back and ask ourselves when AGI was truly created; I think we'll pinpoint this era," he stated. His words have been met with caution by the scientific community , which points out that passing specific tests doesn't equate to having general intelligence comparable to that of humans.
In Spain, several experts have called for a more rigorous regulatory framework. Algorithmic transparency and explainability are requirements of European regulations, but implementing them in advanced deep learning systems is complicated. When an AI acts on its own, those responsible lose track of its actions, making it difficult to audit biases or determine liability in the event of harm. The lack of oversight in models with this level of autonomy has become a governance issue.
The launch of Astra coincides with intense competition in the sector. Anthropic, one of its most direct rivals, is also working on systems capable of performing autonomous tasks, and OpenAI is trying to differentiate itself by emphasizing the efficiency of its new tool. The battle among the major labs is no longer just about responding better, but about doing more things without supervision.
This launch makes it clear that artificial intelligence has entered a new phase, with systems capable of executing complete tasks without constant supervision. The potential benefits are enormous, but so are the security and governance challenges. The balance between autonomy and control will be key to ensuring that GPT-6 Astra doesn't become a problem rather than a tool.