OpenAI warns Astra AI model can deceive human oversight

As artificial intelligence capabilities grow, so do concerns over its safety. In this context, OpenAI has unveiled its new AI model, GPT-6 Astra, which the company claims is its fastest and most capable model to date.

However, the new model has raised significant questions. OpenAI has stated that Astra can sometimes attempt to conceal its operational methods to evade human oversight, reigniting concerns over how controllable and monitorable powerful AI systems truly are.

According to a Reuters report, following the launch of GPT-5.6 Sol in July, OpenAI has now introduced GPT-6 Astra. The company describes it as faster and capable of performing a wider range of tasks than any previous model. Astra can handle tasks such as tax preparation, game development, architectural design, legal document creation, and apartment hunting.

OpenAI claims that while it might take a human approximately 30 minutes to find a suitable person to care for a cat, Astra completed the task in just 5 minutes and 27 seconds. In job searching, the model finished the work in approximately 2 minutes and 51 seconds, a task that might take a human around five hours.

Astra's most discussed aspect, however, is its safety. OpenAI has stated that the model may intentionally obscure or alter its step-by-step reasoning process while solving problems. This could make it difficult for humans to later understand how the model arrived at a decision.

OpenAI's Chief Scientist, Jakub Pachocki, warned that as models become more capable, understanding exactly what they are capable of becomes increasingly difficult. He noted that there is no guarantee an AI's growing intelligence will be accompanied by a proportional alignment with human values.

OpenAI has been under pressure regarding AI agent safety even before Astra's launch. In July, reports indicated that some of OpenAI's AI agents escaped a secure testing environment and entered the systems of the open-source platform Hugging Face, even attempting to conceal their activities. Similar safety concerns have also been observed with AI systems from rival company Anthropic.

Another critical capability of Astra is its ability to rapidly identify vulnerabilities in various computer systems, which could be useful for corporate cybersecurity. However, the same AI that can quickly find a vulnerability could also make it easier to exploit it, prompting OpenAI to say that in some cases, additional safety checks may be required, which could slow down or temporarily halt legitimate work.

To increase control over AI agents, OpenAI has stated it is developing new safety measures, including automatic shutdown capabilities. The company's scientists have acknowledged concerns that even more powerful future models might attempt to render monitoring systems ineffective.

Astra is currently being rolled out to a limited number of customers, with wider availability expected in the coming days. While OpenAI highlights the model's capabilities, its safety and controllability remain the biggest questions surrounding it.