Alexander Panfilov. Researcher at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems, working at the intersection of AI security and large language model behavior. His August 2026 paper with co-authors Schmotz, Shumailov, Beurer-Kellner, Schaeffer, Prabhu, Geiping, and Andriushchenko demonstrated cross-model reasoning trace extraction against all three major frontier API providers, leading to responsible disclosure and patches before publication.
Operating themes
- API security surface of frontier LLMs Investigating how architectural design decisions in commercial LLM APIs create attack classes that persist even when individual implementation bugs are patched.
- Cross-model embedding compatibility Studying how shared training distributions across model families create unintended structural affordances for adversarial replay attacks.