For all matters related to Apollo Research, please reach out to press@apolloresearch.ai.
Marius Hobbhahn
Co-founder & CEO, Apollo Research
I work on AI scheming — AI systems covertly pursuing misaligned goals — and on monitoring and control methods to catch and stop it.
Last updated October 2026
Highlights
2026
U.S. Senate testimony
Testified on misaligned AI at the hearing "Rogue AI: Securing the Homeland Against AI Agent Attacks"
Written testimony →
2026
EU AI Act Scientific Panel
Member of the European Commission's Scientific Panel of Independent Experts on AI
Panel page →
2025
TIME100 AI
Named one of TIME's 100 Most Influential People in AI
TIME profile →
Experience
Apollo Research
2023 – nowCo-founder and CEO; also leading the monitoring team
Epoch
2022 – 2023Part-time research fellow (AI forecasting)
Awards & Engagements
- 2026U.S. Senate — Witness at the hearing on rogue AI agents
- 2026EU AI Act Scientific Panel — Appointed as independent expert
- 2025TIME100 AI — One of TIME's 100 Most Influential People in AI
- 2024Pacific Economic Cooperation Council — AI workshop speaker
- 2024Beijing Academy of AI (BAAI) — Speaker at evaluations panel
- 2023UK AI Safety Summit — Participant at Bletchley Park
- 2022Emergent Ventures — Grantee for AI safety research (15 awardees per cohort)
- 2020German Debate Championships — Best speaker
- 2016German Academic Scholarship Foundation — Scholarship
Selected Press
- 2026Tech Policy Press — Senate Hearing Weighs Threats From Unrestrained AI Agents
- 2025TIME — TIME100 AI profile
- 2024TIME — New Tests Reveal AI's Capacity for Deception
Education
International Max Planck Research School, Tübingen
2020 – 2024PhD with Philipp Hennig — Making Bayesian ML fast and scalable
University of Tübingen
2018 – 2020M.Sc. Machine Learning — Fast Predictive Uncertainty for Classification with Bayesian Deep Networks
University of Tübingen
2016 – 2019B.Sc. Computer Science — Object Detection using the Scattering Transform
University of Tübingen
2015 – 2018B.Sc. Cognitive Science — Inverse Classification using Generative Models
Mentorship & Academic Service
- 2024MATS mentor (Summer 2024) — ML Alignment & Theory Scholars
- 2023SPAR mentor — Mentored 7 AI safety students in mechanistic interpretability
- 2022 – nowReviewer — NeurIPS, ICML and other ML conferences
- 2021Lead organizer — ELLIS doctoral symposium
- 2020 – 2022Supervisor — 3 Bachelor and 1 Master thesis
- 2016 – 2022Teaching assistant — Probabilistic ML, Data Literacy, Math summer course
Skills
- LanguagesEnglish (C2), German (native), French (B2)
- CodingPython; PyTorch > JAX
Selected Publications
- Frontier Models are Capable of In-context Scheming
- Stress Testing Deliberative Alignment for Anti-Scheming Training
- Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
- Large Language Models can Strategically Deceive their Users when Put Under Pressure
- Large Language Models Often Know When They Are Being Evaluated
- Detecting Strategic Deception Using Linear Probes
- Towards Evaluations-based Safety Cases for AI Scheming
- Constitutional Black-Box Monitoring for Scheming in LLM Agents
- Black-box Access is Insufficient for Rigorous AI Audits
- Compute Trends Across Three Eras of Machine Learning
- Position: Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data
Full list on Google Scholar.