Image: generated by METAL AI
Summary
- In a video OpenAI released on September 5, tester Ben Davis shared his experience solving DEF CON's toughest puzzles using GPT-6 Astra
- After receiving hints on a 3x4 Rubik's cube puzzle and a beaded dress puzzle, the model got the answers right, and Davis said one of them marked the world's first solve
- He explained that Astra uses a research-branching structure in which a main agent runs up to 10 sub-agents in parallel, letting it solve problems without losing track
Astra solves a hacker puzzle no one in the world had cracked
In an official video OpenAI released on September 5, tester Ben Davis, who got early access to GPT-6 Astra, shared his first impressions. He said he threw the toughest puzzle challenges from the DEF CON security conference straight at the model, and it solved three puzzles that he and his friends had never been able to crack, plus one puzzle that no one in the world had solved.
OpenAI's official video, "GPT-6 Astra with Ben Davis"
Davis said, "This model was the first in the world to solve this puzzle."
What makes the DEF CON puzzles so hard
DEF CON is a major hacking and security conference held every year in Las Vegas, known for cryptographic puzzle challenges so difficult that participants often spend days on them. Davis said he picked the hardest problems from the conference and handed them straight to Astra. In the end, he said, he and Astra together solved three puzzles that he and his friends had never even made progress on, plus one puzzle nobody in the world had solved before.
Two examples: the Rubik's cube and the beaded dress
He walked through two standout examples. The first was a puzzle that required piecing together several Rubik's cubes arranged in a 3x4 grid to infer a hidden message; after receiving the official hint published by the puzzle's creator, Astra got the answer right on all three attempts. The second involved finding a hidden message across multiple photos of a dress covered in multicolored beads. Despite the poor quality of the photos, Davis said, Astra found the answer once it received a key hint from the puzzle organizers.
| Puzzle | Format | Hint needed | Result |
|---|---|---|---|
| Rubik's cube puzzle | Infer message from cubes arranged in a 3x4 grid | Official hint from the creator | Solved on all 3 attempts |
| Beaded dress puzzle | Extract message from multiple images of multicolored beads | Key hint from the organizers | Solved after hint was given |
Up to 10 agents run at once
Davis explained that when Astra tackles a problem, it first forms a theory, then dispatches a separate agent to test that theory, checks the results, and branches its research from there. The main agent fills up to 10 sub-agent slots and runs them in parallel, keeping only a coordinating role for itself. He said earlier models lacked this intermediate verification step and easily went off track by building on faulty assumptions, but Astra has gotten much better at keeping itself on course. He added that so-called swarm workflows, where multiple agent instances run at the same time, now actually work in practice.
Why this video is coming out now
In a safety evaluation last August, OpenAI said Astra came close to a "critical" cybersecurity risk rating, and on September 1 the company disclosed that the model was the first to receive a "critical" rating under its preparedness framework. The model officially launched on September 3, with OpenAI claiming top scores on long-horizon computer-use tasks and software engineering, alongside announced improvements to alignment. This video followed two days later as a piece of user-testimonial content showing the same model deftly solving real hacking-style puzzles.
To put it plainly, Astra is the same model that OpenAI itself gave a "critical" cybersecurity rating just two days earlier. The puzzle-solving skills on display in this video overlap with the very capabilities that justified that risk rating, so you need to read both stories together to get a real sense of what Astra can actually do.
Editor's take
OpenAI's timing here is telling. Just four days after slapping the model with the heavy label of a "critical" cybersecurity rating, the company showed the same model cracking DEF CON puzzles that hackers take real pride in. Having a risk warning and a capability showcase land this close together reads like OpenAI treating this model's danger and its usefulness as two sides of the same coin. A model this good at security analysis is a weapon for defenders, but it's just as much a tool for attackers, so reading either announcement in isolation only gives you half the picture.
Anyone who has put agent models like this to work in production runs into the same wall every time: running several agents in parallel without an intermediate check, then burning hours building on the wrong assumption. The research-branching design Davis described — forming a theory, verifying it with a separate agent, then branching out from there — looks like it's built specifically to solve that problem. That said, teams considering deployment should keep an eye on the cap of 10 slots. Breaking a task into smaller parallel pieces also drives up token usage and cost, so it makes sense to pilot this first on work with heavy verification needs, like security analysis or reverse engineering, and measure the cost-benefit before rolling it out further.
More videos like this, showing real-world use cases, will likely follow in the coming weeks. At the same time, the security community is likely to keep testing and trying to reproduce Astra's cyber capabilities, and those results will become the next benchmark for whether OpenAI's risk rating was justified.
Comments