GPT-6 Astra cracked a DEF CON puzzle

By OpenAI

Categories: AI, Product

Summary

GPT-6 Astra solved complex DEF CON puzzles by orchestrating 10 parallel agent slots—demonstrating that explicit instructions combined with multi-agent collaboration enables AI to tackle problems humans spend days solving.

Key Takeaways

  1. Astra solved a 3x4 Rubik's cube puzzle pattern 3/3 times with official hints, showing reliable performance on complex spatial reasoning tasks when given proper guidance.
  2. The model uses a main orchestrating agent with 10 parallel agent slots running simultaneously to test theories and execute subtasks—enabling distributed problem-solving across multiple threads.
  3. AI performance on hard problems scales dramatically with explicit instruction on workflow and orchestration patterns—moving from trial-and-error to structured hypothesis-testing loops.
  4. Astra demonstrates hypothesis-driven problem-solving: generates theory, dispatches agents to test, observes results, iterates—mimicking human debugging workflows at scale.
  5. Complex puzzle-solving improved from unexpectedly failing to consistently succeeding—the critical variable was receiving the same hints humans used, indicating context parity drives AI breakthrough performance.

Related topics

Transcript Excerpt

I'm very very impressed. I went to a big conference called Defcon where there are these big puzzle challenges that me and my friends spent days trying to solve. I took some of the hardest problems from that challenge and I threw them at the new model Astra and it has gotten two of them that I did not expect it to get. It solved this one big puzzle which was like a bunch of Rubik's cubes that were arranged in this weird 3x4 pattern that you had to deduce a message from. It was able to get that three out of three times. The one thing it did need is it got the official hint from the puzzle creators. So it's the same thing we got when we were doing it. But once you gave it that, it got the actual solution. It is very intelligent. It is very good at grinding down very complex tasks. The model c…

More from OpenAI