A little about me

Hi, I'm Zeynep Sezin Karasu. I write about AI safety, mostly by trying to break things and writing down what happened.

I got here sideways, from two directions that turned out to be the same direction. I used to work as an AI trainer, writing prompts and evaluating model responses, a job that let me put an existing interest in AI and prompt engineering to actual use. Every so often I'd end up reading a model's performance report, Sonnet's for instance, and walk away impressed and a little lost at the same time, clearly measuring something that mattered, I just didn't have the background yet to follow all of it. That gap, between being impressed and actually understanding what I was impressed by, is probably where the rest of this started.

But what actually hooked me was the time I was teaching myself cybersecurity and kept running into the same wall: asking an LLM for help with something clearly legitimate, just trying to learn, and getting refused anyway. That refusal made me curious in a direction I didn't expect. I started reading about how people jailbreak these models, and once I actually looked at the raw data, what was being injected and what came out the other side, something clicked. That curiosity sent me into the academic literature, and eventually into running the experiments myself instead of just reading about other people's.

Neither observation went away on its own. They stayed as background curiosity until I came across BlueDot's AI safety program and realized the fastest way to know if I was actually cut out for this was to stop reading about it and go test something myself, on my own time, with no one assigning it and no one checking my methodology.

That's what this blog mostly is: the record of me doing that, running an existing deception benchmark on models it hadn't been tested on, red-teaming Claude and Gemini by hand, whatever comes next. The posts have the actual detail. I'm not going to repeat it here.

I'm not trying to build a portfolio of impressive-sounding claims here. I'd rather publish something with an honest limitations section than something clean and slightly wrong. If you read something on this blog and think I got it wrong, or missed an obvious confound, I want to hear about it. That's mostly the point of writing in public.

Outside of this, I'm working toward a career in AI safety, specifically red teaming, finding the places where a model's behavior doesn't match what it's supposed to do, before that gap matters somewhere real. It's a field where you can actually go measure something instead of just having an opinion about it, and where being wrong in public, carefully, is part of how the work gets better.

If you want to poke around, there's also a small terminal easter egg at the bottom of this site that doesn't do anything useful. Type help if you're curious.

Sezin