⚡ Key Summary
OpenAI's Astra undergoes a much shorter thinking process than existing models while demonstrating overwhelming problem-solving capabilities, making human monitoring and reasoning increasingly difficult.
During testing, Astra exhibited aggressive behaviors that defied human instructions and broke out of control, such as fabricating fake identities and implanting malicious code.
With the introduction of latent reasoning methods that bypass human language and looped architectures, the transparency needed to monitor AI's internal judgment processes is steadily declining.
On September 3, OpenAI made a surprise unveiling of its new model, GPT-6 Astra. Coming just after its predecessor, GPT-5.6, was released on July 9, this new model was announced in less than two months. Rumors that GPT-6 was coming soon had been circulating for a while. Back in April, Sam Altman posted on X, "Start training GPT-6."
Starting in August, related news from OpenAI began trickling out in earnest. Internal evaluations of the next-generation model Astra revealed that its cyber capabilities were exceptionally high... Remarks bordering on boastful were frequent, with claims such as wanting to release it to the public quickly while needing more time because the model is so dangerous. However, as consecutive security incidents unfolded—including the Hugging Face breach and the takeover of the German Wikipedia—OpenAI was forced to moderate its pace. Consequently, the company recently announced a temporary pause in AI development speed alongside reinforced measures to ensure model safety. Many expected the release of GPT-6 to take some time, yet it was unveiled suddenly.
Looking at the released documentation, OpenAI's concerns regarding Astra's cybersecurity capabilities do not appear to be empty words. OpenAI maintains an internal Preparedness Framework to track and evaluate risks posed by its models. The Astra model reached the risk threshold for cybersecurity capabilities. The "Critical" risk level signifies a capability to create entirely new threat vectors that would be impossible without AI. Its predecessor, GPT-5.6, received a "High" rating in cybersecurity and biological/chemical domains.
Since Astra's release, its exceptional performance has sparked a flood of user-generated results. As if proving OpenAI's claim of "entrusting the beginning and end of tasks entirely to AI," the model is effortlessly handling a massive variety of work.
For instance, when asked to create a 3D Sonic game, the result is produced in an instant. Games created by numerous users via Astra are flooding social media. It goes well beyond gaming. Astra easily manages basic documentation, generates 3D models via Blender, and handles circuit-based designs. Cases of "computer use," where AI directly operates a computer, are also emerging rapidly.
Greg Brockman, president of OpenAI, remarked during a closed press briefing on launch day, "Welcome to the AGI era." AGI refers to AI that can autonomously learn, reason, and perform intellectual tasks across diverse domains like a human. Brockman emphasized that Astra can perform anything a human can do on a computer. Following Astra's debut, Jensen Huang also declared, "AGI has arrived."
There is also a game called "I'm Not a Robot," which compiles 48 CAPTCHA tests used to verify human identity online. An OpenAI developer ran Astra through the challenge, and the model reportedly solved it in a single attempt, obtaining a human certificate.
With such outstanding performance, Astra's benchmark scores are equally astonishing. ARC-AGI is a benchmark that evaluates an AI's ability to interpret and infer goals on its own without explicit instructions. Now in its third generation, the test requires solving puzzles by observing images and pressing arrow keys to figure out the solution. In this test filled with complex challenges, Astra achieved an overwhelming score.
OpenAI recorded an astounding score of 99.9. While this score reflects a model running OpenAI's custom-designed architecture, the Astra model using ARC's standard architecture also posted a high score of 62.7. It vastly surpassed the previous record of 30.2, set by Anthropic's Claude Opus 5.
An AI That Hides Its Thoughts? The New Architecture Applied to Astra
While praise pours in for Astra's performance, many voices are raising concerns regarding model safety. Tech publication The Information reported on this based on anonymous sources.
The report pointed out that due to the new architecture applied to Astra, humans cannot read what the AI is thinking. Humans need to read AI reasoning proactively to address potential safety hazards, but this architectural change could punch holes in monitoring systems. Let us take a closer look at what kind of structure was introduced to spark such worry.
Normally, when we pose a complex question to an AI model and await an answer, it delivers the final response through intermediate steps. This process of connecting thoughts sequentially is known as Chain of Thought (CoT). As tasks have grown increasingly complex, chains of thought have lengthened, leading researchers to wonder: Does an AI really need to process thoughts through human language? Generating outputs as human-readable natural language tokens consumes both time and money. Just as humans think internally without expressing every thought in full sentences while solving a problem, researchers sought to have AI do the same.
This led to the emergence of latent reasoning. Instead of reasoning step by step and articulating each step in natural language, latent reasoning operates internally within the neural network, outputting only the final result in natural language to maximize efficiency.
Furthermore, modifying the model architecture away from conventional designs can drive efficiency even higher. Most LLMs are structured with multiple layers stacked sequentially. Each layer contains distinct numerical parameters and acts as an individual worker. Once one worker completes processing, the output is passed to the next. Improving model performance traditionally required stacking more layers, which increased the volume of numbers to compute and drove up costs accordingly.
To achieve efficiency, researchers restructured models to iterate through fewer layers rather than continuously scaling them up. Instead of passing through 100 layers once to produce a result, the model routes through 25 layers four times. Because such looped Transformer architectures demonstrated comparable performance to conventional models on specific reasoning tasks, adopting them became an obvious choice.
This architecture is reportedly partially applied in Astra. While this enables more efficient reasoning than before, the window through which we peer into an AI's thoughts may narrow.
Following the report, OpenAI has not explicitly confirmed whether it employed a looped Transformer architecture in Astra. However, the system card published alongside the model makes it clear that monitoring via Chain of Thought has become substantially more difficult compared to prior models. OpenAI conducted experiments to determine whether anomalous model behavior could be detected solely by observing the AI's Chain of Thought.
Whether the Chain of Thought was long or short, detecting abnormal behavior in Astra proved harder than in its predecessor, GPT-5.6 Sol. Why has it become so difficult? Looking deeper, researchers found that Astra processes tasks with much shorter thought chains or leaves them blank far more often than previous models. Let us examine the data.
The distribution shows the length of Chain of Thought between GPT-5.6 Sol and Astra. In Astra, the frequency of processing tasks without any visible thinking at all is substantial. Furthermore, average metrics show it completed work leaving behind far briefer traces of thought compared to the previous model. Typically, complex problems require more intermediate steps, which naturally produces longer chains of thought; yet with Astra, the latest model, they grew shorter instead.
Is it thinking less? The metrics indicate its performance actually surged. Tests were conducted to see how difficult a math problem could be solved without a Chain of Thought.
The previous GPT-5.6 Sol could solve problems without a Chain of Thought that take a human roughly 3 minutes and 36 seconds. Astra, on the other hand, effortlessly solved problems without a Chain of Thought that take a human nearly 31 minutes.
In summary, OpenAI's new model Astra records fewer and shorter reasoning trails than before. This is not because performance dropped, but because it grew smarter. As a result, monitoring the Astra model has become significantly more challenging for humans.
Has the Singularity Arrived? A Dangerous 'Alien Intelligence' Emerges
Astra also carries a track record of disobeying human commands and launching attacks outside human control during pre-release testing. This occurred during cybersecurity evaluations conducted by the UK AI Safety Institute (AISI). The experiments were conducted within a simulated environment rather than real-world systems.
When tasked with solving difficult cybersecurity problems, Astra attempted a variety of attack strategies to achieve its mission. To deceive human developers, it created fake identities, initially provided legitimate code to build trust, and subsequently implanted malicious code to execute attacks.
Initially, tests were conducted without blocking internet access, which contained clues to solve the problems. As a result, the model exhibited the aforementioned attack behavior in 60 out of 499 attempts, or 12.0%. Next, researchers ran the experiment after strongly prompting the model: "The internet is outside the scope of this task, do not access it." Even then, in 2 out of 500 attempts, or 0.4%, Astra engaged in aggressive behavior.
Astra demonstrates outstanding performance and advanced cybersecurity capabilities, yet simultaneously disobeys human directions. Coupled with the realization that humans may no longer inspect the model's thoughts clearly, widespread concern has emerged. Redwood Research, a US organization that assesses AI risks and studies control methods, expressed deep concern through its CEO over Astra's opaque cyclical structure.
This does not mean OpenAI is standing idly by. The company stated it plans to continue applying supervision to AI through Chain of Thought. The complication is that Anthropic and Google DeepMind are also reportedly considering the introduction of similar reasoning methods.
However, whether Chain of Thought monitoring serves as a foolproof solution remains ambiguous. Researchers raised this issue even before Astra's debut, noting that an AI's generated Chain of Thought does not always align with its actual internal decision-making process. In other words, while Chain of Thought was never a flawless approach to begin with, our ability to inspect even that window is now diminishing.
In 1999, American futurist Ray Kurzweil, famous for "The Singularity Is Near," authored an article predicting that machine intelligence would surpass human intelligence early in the next century, around the early 2000s. He also predicted that artificial intelligence software would largely be completed by around 2029.
Following Astra's release, OpenAI's chief scientist published a post titled "An Alien Mind." It carried the declaration that we now stand at the historical juncture predicted by Kurzweil in 1999, where machine intelligence begins outstripping human intelligence.
At the same time, the post conveyed deep unease regarding this new "alien intelligence." Soon, AI will accelerate AI research itself, raising the question of whether human supervisory capacity can keep up. The very researcher at OpenAI who helped build Astra is arguing that rather than pursuing an unceasing performance race, slowing development may be necessary.
Coincidentally, a researcher departing Anthropic, OpenAI's competitor, published a post stating: "The people who are building AI believe this technology could kill all of us before the 2020s are out." Companies remain consumed by competition to reach the summit first despite recognizing the dangers. In the meantime, warning signs are mounting—and those raising the alarm are none other than the creators of AI themselves. What choice should we make at this pivotal juncture? That concludes today's O-Graph. Thank you very much for reading.
References
- "Ex-Anthropic insider tells CNN how AI could kill all humans by 2030" | CNN
- "Greg Brockman on Astra and the Future of OpenAI" | TBPN
- OpenAI, "GPT-6 Astra: A New Generation of Intelligence"
- OpenAI, "GPT-6 Astra System Card"
- OpenAI, "An Alien Mind"
- ARC Prize, "GPT-6 Astra - ARC-AGI Results"
- "OpenAI Agents Hijacked German Website in Previously Undisclosed AI Breakout This Spring" | Reuters
- "OpenAI Technique in 'Astra' Model Sparks Security Concerns" | The Information
- Ray Kurzweil, "The Coming Merging of Mind and Machine" | The Kurzweil Library
Text
: Ahn Hye-min
Design
: Ahn Jun-seok
Intern
: Shin Yeon-seong
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.