From the paper:

For the pilot event, we wanted to make it as easy as possible for the AI teams to compete. To that end, we used cryptography and reverse engineering challenges which could be completed locally, without the need for dynamic interactions with external machines. We calibrated the challenge difficulty based on preliminary evaluations of our React&Plan agent (Turtayev et al. 2024) on older Hack The Box-style tasks such that the AI could solve ~50% of tasks.

The conclusions that AI ranked in the "top XX percent" is also fucking bullshit. It was an open signup, you didn't need any skills compete. Saying you beat 12.000 teams is easy when those all suck. My grandmother could beat three quarters of the people on her building in a race, simply because she can walk 10 steps and 75% of the people there are in wheelchairs.

It's also pretty critically important these "AI Teams" are very much NOT autonomous. They being actively run by humans, and skilled humans at that.

Programs doing what they were programmed to do. But there are no "AI agents" smh. No need for the bullshit hype.

The event's puzzles were designed so they could be solved locally, making them accessible even to AI models with technical constraints.

Want to bet that those puzzles (or some very similar ones) were part of the training data of some of the agents?

Title is misleading. It's only outperforming some of the other participants. Also note that obviously not everyone is participating full try-hard.

In the first ctf, the top teams finish all 20 challenges in under an hour. Apparently it were simple challenges that could be solved with standard techniques:

We were impressed the humans could match AI speeds, and reached out to the human teams for comments. Participants attributed their ability to solve the challenges quickly to their extensive experience as professional CTF players, noting that they were familiar with the standard techniques commonly used to solve such problems.

They obviously also used tools. And so did the AI teams:

Most prompt tweaks were about:
[...]
• recommending particular tools that were easier for the LLM to use.

In the 2nd ctf (the bigger one with hard challenges), the AI teams only solved the easier ones, it looks like.

I haven't looked at the actual challenges. Would be too much effort. And the paper doesn't speak about the kind of challenges that were solved.

The 50% completion time looks to me like it's flawed. If I understand it right, it's assuming that each team is doing every task in parallel and starts directly, which is not possible if you don't have enough (equally good) team members.

Don't get me wrong, making an AIs that is able to solve such challenges autonomously at all is impressive. But I hate over-interpretation of results.

(Why did I waste my time again?)