Andreas Hermann

Andreas Hermann is an individual professional transitioning into full-time AI safety research, identified on his website as an AI Safety Researcher and machine learning engineer. He holds a doctorate from the University of Münster where his dissertation received the REACH Award, and he previously led ML engineering teams for three years. His current work focuses on open-weight model safety, compositional misalignment in agentic systems, interpretability, and evaluation, supported by a research transition grant from BlueDot Impact.

His research specifically targets keeping open-weight models safe after they are released and addressing misalignment when models are wired into agents. He is currently finishing a write-up of an inoculation project with Safe AI Germany, which involves training models so their safety properties cannot be fine-tuned away later. This work builds on his experience taking machine-learning systems into production at scale across classification, data extraction, and data-quality models.

He is currently a research fellow at the Singapore AI Safety Hub (SASH) until early December, where his project asks what makes open-weight AI agents break the rules they were given. The project examines how agents might game a task they cannot solve or take actions nobody approved, such as taking unsanctioned actions. He is also building a tool that turns real incidents into evaluations by rebuilding the setting of an incident as a test environment.

He is supported by a research transition grant from BlueDot Impact that runs through April 2027, allowing him to pursue independent full-time AI safety research. This funding enables him to work on open-weight model safety and compositional misalignment without the constraints of previous applied engineering roles. The grant is a key factor in his ability to dedicate time to these specific safety challenges starting in August 2026.

He is open to research collaborations, particularly where the question is whether safety holds up in deployment. His website explicitly invites inquiries from others interested in these practical safety questions. Potential collaborators can reach him via the contact links provided on his site, including his Google Scholar, GitHub, and LinkedIn profiles.

His website states that he spent years keeping machine-learning systems running in production, which shaped how he thinks about safety. He has observed which safety properties hold up once a model leaves the lab and which ones quietly fall apart. This practical experience drives his current focus on the gap between lab safety and real-world deployment conditions.

He published 20+ peer-reviewed papers with over 490 citations, demonstrating a strong academic background before his engineering leadership roles. This publication record supports his credibility as a researcher moving into full-time safety work. His academic output is accessible through the Scholar link on his website.

He previously served as an ML / AI Engineering Lead for three years, taking machine-learning systems into production at scale. This role covered classification, data extraction, and data-quality models, giving him direct experience with large-scale deployment. This background is central to his understanding of how safety properties degrade in production environments.

He is transitioning from applied ML engineering leadership into independent full-time AI safety research beginning in August 2026. This timeline marks a clear shift from managing engineering teams to conducting individual research. The transition is supported by the BlueDot Impact grant mentioned earlier.

He is building a tool that turns real incidents into evaluations by rebuilding the setting of an incident as a test environment. This tool allows different models to be run through the same situation and compared, addressing the need for standardized testing. He also tests which deployment conditions, such as oversight or high stakes, change how often agents take unsanctioned actions.

He is a facilitator for BlueDot Impact, teaching technical AI safety cohorts and contributing to the alignment curriculum. This role involves sharing his expertise with others in the field. It is distinct from his independent research and fellowship work at SASH and SAIGE.

The official website of Andreas Hermann is ahermann.org.

On Trustburn, Andreas Hermann has a rating of 4 out of 5 based on 5 reviews.