Anthropic Fellows Program opens May and July 2026 cohorts across six tracks — scope statements signal where lab safety research moves next
Anthropic's expanded Fellows Program runs both a May 2026 and a July 2026 cohort, with detailed scope statements published across six tracks: scalable oversight, adversarial robustness, AI control, model organisms, mechanistic interpretability, and model welfare. The track-by-track scope reads as the most precise public signal of which open problems Anthropic has decided to invest external-researcher capacity against.
The two-cohort cadence (May and July starts) is the operational news. Through 2024-2025 the lab ran roughly one Fellows cohort per year; the 2026 expansion to two cohorts per year doubles the throughput and signals that the Fellows pipeline has moved from experimental program to load-bearing institutional infrastructure. Combined with OpenAI's parallel Safety Fellowship, the alignment talent pipeline is now structured as parallel lab-funded programs rather than as a single-lab differentiator.
The scope detail across the six tracks is the methodological signal. Scalable oversight focuses on bootstrapping supervision when humans cannot verify model output. Adversarial robustness emphasizes red-team-resistant evaluation. AI control treats containment as a methodological track equal to alignment. Model organisms ports the biological-research "organism" pattern to ML — using small purpose-built models as objects of careful study. Mechanistic interpretability has its own track separate from broader interpretability, signaling the lab's bet on direct-circuit-reading methodology. Model welfare is the most institutionally novel track and the one that most clearly signals Anthropic taking positions other labs have not yet. Six tracks, six methodological bets, all funded as parallel investments.
Anthropic Alignment — Fellows Program 2026 May and July cohorts → · Anthropic Alignment — Automated Weak-to-Strong Researcher → · Claude 5 Hub — AI Safety 2026 Alignment Research Breakthroughs →