AI Safety at the Frontier

AI Safety at the Frontier is a free monthly newsletter published on Substack by Johannes Gasteiger, a researcher working on Alignment Science at Anthropic. Gasteiger previously worked at Google Research, TU Munich, DeepMind, and Meta AI, and completed his PhD at TU Munich in 2022 with a focus on machine learning for graphs and molecules. Founded on July 18, 2024, the newsletter's stated mission is to surface the best AI safety research on frontier models. Each month, Gasteiger reviews more than 50 papers and curates a subjective selection of the most interesting and important ones, providing summaries and context. The publication operates entirely as a personal project alongside his work at Anthropic, with no paid subscription tier, no team, and no formal organizational structure. Content focuses on ML-oriented AI safety research, including topics such as reward hacking, AI scheming, sandbagging detection, capability elicitation, model auditing, and alignment and control mechanisms. Issues are cross-posted to LessWrong, where they reach additional readers in the AI safety community. The newsletter had approximately several hundred subscribers as of early 2026.