Insider Training: Data and Reward Poisoning Attacks with Data-Level Defences

I’m researching data poisoning attacks on frontier models through RL (GRPO) fine-tuning and SFT experiments, and developing blue-team evaluations such as InspectAI evals, statistical dataset analysis, and controllable oversight methods.

This work is supervised by Mary Phuong (Google DeepMind). The project has received $120K+ in extended funding and will be submitted to a top AI conference (ICML 2026).

A preliminary preprint is available below for anyone interested. Please note that the work will change substantially before final submission.