Spectator Sport

Day 5 of Inkhaven: 30 Days of Posts

1... 1... 1. The finger tapped the keyboard and lines of code and commentary began to stream onto the screen. The AI methodically iterated to surface various new vulnerabilities in the system. It explored various vectors of attack from database injection attacks, common shell vulnerabilities and overflow attacks. The participant took off his university jersey. He looked at the documentation for the competition, copied it into Claude's context, and gave it the rules of the competition before switching over to twitter. Claude went about its work. Lines of code appeared, were tested, failed, and more lines flowed out to replace them. The student scanned the new outputs with some interest and then tabbed away. The AI had the quiet relentlessness of a dishwasher moving through thousands of cycles.

The competition went on around him, promising young minds from across the country who'd gathered here because they were excited about cybersecurity. The organisers had spent weeks in the leadup to the competition trying to determine a reasonable LLM policy. I'm not sure what the final policy was, but ultimately that question had consumed most of their effort. Maybe that's what participants should have been assessed on.

I'd been brought in as a judge for this whole thing, but so far, hadn't found much human activity that warranted judging.

Claude sputtered to a pause as its first attempts failed to crack the problem. The student read the output, asked for some suggestions. Thinking for a moment he selected one of the options. That didn't work. "Keep trying, you got this!" came the response from the participant. I wondered how Claude felt about all this.

I idly wondered why they hadn't completely automated their role in the process yet. A simple hook that simply responded with "You can do it!" should do the trick. You could set it up in a few minutes. You could probably ask Claude to write it for you.

At a different table, someone else seemed to have found a much more promising approach. She looked about 14. I wasn't sure if she was a participant or someone's child, but her team's score was near the top of the board.

The hours passed. Cries of success rang out occasionally, though from where I sat it was hard to tell whether they were celebrating something they'd done or something they'd witnessed.

As the competition was wrapping up, I happened to walk past the student from the beginning. He'd stopped working on the problem, and hadn't been able to get to the end. I overheard him chatting to the organiser next to him about the competition. He mentioned he'd tried a few different approaches, that the authentication layer had seemed promising for a while but ultimately it was a dead end. He asked about their summer internship programme. He said he'd learned a lot and really enjoyed having to work under a tight deadline. The organiser nodded, said they'd send him a link where he could apply. He shook her hand and grabbed his bag.