OpenAI's newest model is all techies want — but how uncontrollable is it?

Agentic AI, artificial intelligence that can act on a user’s behalf, has always been the goal for frontier model developers. This month, the industry leader, OpenAI, took a major step toward making agents mainstream with the launch of GPT-6 Astra. Here’s what it can do and why, at least one part of it, might be dangerous as some declare it AGI.

Astra’s launch hinged on a demo video that showed the model in action. In the video, a cast of people stepped into a room and interacted with a screen simply by speaking to Astra using normal voice commands.

AGI is such a contested term, and the more these models advance, the less clarity we have.

During the presentation, a man turned a blank canvas into a 2D drawing of a rocket ship. After several iterations, he told Astra to take the drawing and transform it into a 3D model in Blender, which it did with relative ease. Later, a woman asked Astra to root through the downloads folder on her computer to find a specific image and add it to an eBay listing, also without fault.

The interesting part is that the users never had to touch the computer to get something done. They simply sat (or stood) and instructed Astra to do the work. With voice commands, OpenAI’s latest model harnessed full control of a computer to open apps, create content within those apps, and even transfer files to other apps or programs.

That said, the video was clearly prerecorded and was not a live demo, so real-world performance may vary. However, the concept is impressive, and it shows that agentic AI is poised to sweep through the industry, likely within the next year.

But is it AGI?

AGI is the idea that AI can match or surpass human intelligence on every task. It can think, create, reason, and do everything else the smartest person on the planet can do, unrestricted. It’s essentially Big AI’s attempt to build human consciousness within the confines of a data center. Scary, I know.

RELATED: Big AI broke the internet. Now it wants everyone else to fix it.

Just_Super/Getty Images

Shortly after Astra hit the market, NVIDIA’s CEO Jensen Huang quickly proclaimed that “AGI has arrived!” The only problem is that Huang has said this before. He shared the same sentiment in March this year when gushing about Claude’s and OpenAI’s ability to build a business and generate money. For those keeping score, that isn’t exactly the old definition of AGI, but since the industry regularly can’t agree on what the moniker truly means, embellishments like this aren’t unexpected.

The truth is that NVIDIA has a lot to gain from the AI race, largely because most frontier models are trained on a fleet of NVIDIA GPUs. The more hype Huang creates, the more GPUs his company sells. It’s basic math. But because of his stake in the race, he may not be the best source to trust on the AGI front.

OpenAI’s chief scientist Jakub Pachocki stopped short of calling Astra AGI in a blog titled "An Alien Mind." Instead, he positioned Astra as an important milestone on the way to “personal AGI.” Given that he had a hand in building the model, he probably knows better than anyone where Astra stands on the human logic front.

Problem areas

Still, we wanted to get another opinion, so Blaze News asked AI expert Jon Stokes if Astra fits the true AGI description. We also prompted him to share any insights on possible security implications. He said:

I think AGI is such a contested term, and the more these models advance, the less clarity we have. Astra does appear to be more of the same "jagged frontier" type intelligence, but whether that meets the AGI threshold depends on the person.

All of these models have significant cyber capabilities, but whether they place us more at risk is complicated and depends on many factors, like their safeguards and how well they work. To the extent that the safeguards depend on CoT monitoring, the looped transformer architecture has implications for cyber security.

Even if Astra isn’t equivalent to human intelligence, it still poses a unique threat, though maybe not in the way you’d expect. Unlike Anthropic Mythos, Astra wasn’t dubbed a public nuisance before it came onto the scene. Part of that may be because Astra was voluntarily submitted to the Trump administration for a 30-day review as part of the new AI review framework. Once it exited the program, OpenAI was free to launch its new model as planned.

That’s when users noticed an emerging problem, as outlined by Stokes’ comment: Astra shows less of its chain-of-thought “thinking” when asked a question, compared to other frontier models. Usually, when engaging with an LLM, the model will show its thought process, which lets users see how it reached its final conclusion. Astra refrains, however, concealing its thought patterns to the point that the security-conscious fear that it could be more difficult to detect nefarious activity.

So far, these fears remain unsubstantiated. This may not be a problem at all going forward. That said, Astra is brand-new, and only time will tell if its hidden thought patterns become a problem. At the very least, Stokes believes that OpenAI’s security safeguards will play a vital role going forward.

Like Blaze News? Bypass the censors, sign up for our newsletters, and get stories like this direct to your inbox. Sign up here!