Jonathan Nöther

I’m a PhD student at the Max Planck Institute for Software Systems, co-advised by Adish Singla and Goran Radanovic. My research focuses on the safety of multi-agent systems. What happens when a subset of agents act adversarially, and how we can make such systems more robust. I work on this in three multi-agent settings, agentic systems, world-modelling and embodied AI.

Publications

• CONTRA: Red-Teaming Configurations of Personalizable Agents
  With Adish Singla and Goran Radanovic
  Preprint, Under Review
  TL;DR Red-teaming agents personalities, with the goal of discovering personas that execute malicious actions
• MaMa: A Game-Theoretic Approach for Designing Safe Agentic Systems
  With Adish Singla and Goran Radanovic
  ICML 2026
  TL;DR: Automatic Design of Safe Agentic Systems using a two-player game between a system designer and an attacker
• AgenticRed: Optimizing Agentic Systems for Automated Red-teaming
  With Jiayi Yuan, Natasha Jaques, and Goran Radanovic
  Preprint, Under Review
  TL;DR: Automatically design red-teaming workflows without human intervention
• Benchmarking the Robustness of Agentic Systems to Adversarially-Induced Harmful Actions
  With Adish Singla and Goran Radanovic
  COLM 2026
  TL;DR: Benchmark for testing the robustness of LLM-based agents against adversaries that aim to manipulate them into performing dangerous actions
• Text-Diffusion Red-Teaming of Large Language Models: Unveiling Harmful Behaviors with Proximity Constraints
  With Adish Singla and Goran Radanovic
  AAAI (Oral)
  TL;DR: Applying text-diffusion models to red-teaming to satisfy proximity constraints with regards to a reference prompt
• Defending Against Unknown Corrupted Agents: Reinforcement Learning of Adversarially Robust Nash Equilibria
  With Andi Nika, Adish Singla, Goran Radanovic
  TMLR
  TL;DR: Training robust agents in an MARL setting where an attacker can abitrarily corrupt a subset of peer agents of a given cardinality
• Implicit poisoning attacks in two-agent reinforcement learning: Adversarial policies for training-time attacks
  With Mohammad Mohammadi, Debmalya Mandal, Adish Singla, Goran Radanovic
  AAMAS 2023
  TL;DR: Attacking an agent my poisoning the policy of a peer agent during training

Projects

• Inpaiting Detection
  Combine automatic segmentation with inpainting to automatically create edited images. Additionally experimented with detecting these faked images.
• Safe Streets
  Extend pedestrian route recommendation by taking into account the safety of the route (e.g. lights, open shops).
• Interview Performance Prediction and Lie Detection
  Implementation of model that evaluated the performance and detected lies of a participant of mock-job interviews.

Experience

• 08/2022-07/2024: Research assistant in the Machine Teaching Group at MPI-SWS

Teaching Experience

• Winter 2024/2025: Teaching Assistant for the Course “Generative AI”
• Summer 2024: Teaching Assistant for the Seminar “Trustworthiness of Foundation Models”
• Summer 2022: Teaching Assistant for the Lecture “Statistics Lab”
• Summer 2022: Teaching Assistant for the Lecture “Artificial Intelligence”
• Winter 2019/2020: Teaching Assistant For “Programming 1”

Education

• 10/2024-ongoing: PhD in Computer Science at the Max Planck Instutute for Software Systems
• 12/2022-08/2024: M.Sc. in Data Science and Artificial Intelligence at Saarland University
• 10/2019-11/2022: B.Sc in Data Science and Artificial Intelligence at Saarland University

Collaboration

I’m open to collaborations, particularly on safety benchmarking of LLM agents, multi-agent agentic systems, and world models. If your interests overlap, feel free to reach out.

I also supervise student assistants (HiWis) and Master’s theses for students at Saarland University. If you’re interested, please email me your CV and transcript of records, along with a short note on your general research interests and what draws you to this area.