Samuel Nellessen
AI Research Engineer, LLM Security & Safety
📍 Berlin, Germany
About Me
I’m an AI security researcher building systems that autonomously find failure modes in LLM agents. I am currently an independent Foresight AI Safety grantee at the Foresight Node in Berlin and a student researcher in Tal Kachman’s lab at Radboud University.
My current work spans long-horizon latent adversarial training, automated agent-to-agent jailbreaking, and adversarial attacks on LLMs. Recent projects include Slingshot and Tag-Along Attacks, multi-GPU CISPO training on Slurm/HPC, and mechanistic interpretability work on refusal behavior.
Previously, my focus was computational neuroscience and philosophy. I studied AI at Radboud University. In 2024, as a Foresight Neurotech Fellow, I worked with Roshan Cools to build Bayesian models of controllability in clinical depression. After completing the ARENA v5 bootcamp I transitioned fully into empirical AI security. I also contribute to AI safety tooling, including UK AISI’s inspect_ai and Prime Intellect Environments.
When I step away from the desk, I try to stay entirely offline. I spend my time training calisthenics, singing in classical choirs, and wrenching on vintage bicycles. My current pride is a Koga Miyata 1989 Carbontech 5000 that I rebuilt into a time trial bike with a vintage Campagnolo Record groupset. I also write autofiction under a pseudonym. If you manage to dox my pen name, the first round of coffee in Berlin is on me. If you have particularly strong search skills, you might even unearth an ancient YouTube channel of me playing guitar.
I am always happy to chat. Book a 1-on-1 or email: samuelgerrit.nellessen{at}gmail.com