The Anthropic Fellows Program is a funded research program in which people who are not Anthropic employees work full-time on AI safety research with Anthropic mentors. This Alignment Science Blog post announced it as "a pilot initiative designed to accelerate AI safety research and foster research talent". The blog's index lists the post under December 2024; a crosspost of the same text on LessWrong is timestamped 30 November 2024 (23:47 UTC), and the first Wayback Machine capture of the blog page is from 2 December 2024. The post reports no research results. A notice at the top, dated July 2026, points to the later announcement; this note follows the text retrieved on 13 November 2026.
What the post announced
The post describes a single cohort of "10-15 Fellows" working full-time for six months, matched with Anthropic mentors. Its stated reason is that Anthropic's researchers had "completed about a dozen projects with external collaborators" in the previous year, many published "at top venues", and that a formal program would widen the pool of researchers able to move into safety work. Footnote 1 names three such collaborations: debate with more persuasive models (arXiv:2402.06782), unfaithful chain-of-thought explanations (arXiv:2305.04388), and vision-language models as zero-shot reward models (arXiv:2310.12921). The program's aim, in the post's words: "we aim for every Fellow to have produced a (co-)first authored research paper in AI safety."
Terms as stated
The section "What Fellows should expect" sets out the terms:
- Status. Fellows "will not be officially employed by Anthropic".
- Stipend. A weekly stipend of $2,100 with access to benefits.
- Research budget. Fellows "will collectively have access to a budget of approximately $10,000 per fellow per month" for compute, human data and other research costs.
- Models. "All research projects will be conducted using publicly available APIs and open-source models."
- Hours. Full-time, which the post defines as 40 hours per week, with other obligations such as coursework allowed if that commitment holds.
- Place. Shared workspaces in the Bay Area and London (not Anthropic offices), with visits to Anthropic's San Francisco or London offices.
- Timeline. Project selection and mentor matching in March 2025, onboarding in mid-March 2025; later cohorts, "if we have subsequent cohorts", on a rolling basis.
- Deadline. Applications for the first cohort by 20 January 2025, reviewed on a rolling basis.
- Employment. The post states that it does not guarantee full-time offers.
Eligibility and selection
The post says the program is "focused on helping mid-career technical professionals transition into the AI safety research space" but is open to applicants at any career stage. Listed expectations include strong Python, a technical background in computer science or machine learning, and US or UK work authorization with the ability to work from Berkeley or London (other locations case by case). Selection criteria are grouped as "Technical Excellence" and "Research Potential". The post encourages applications from people new to the field and from groups underrepresented in tech.
Research areas and mentors
The post lists 13 potential mentors: Ethan Perez, Jan Leike, Jascha Sohl-Dickstein, Andi Peng, Samuel Marks, Joe Benton, Akbir Khan, Fabien Roger, Alex Tamkin, Kyle Fish, Nina Panickssery, Mrinank Sharma and Evan Hubinger. It names five research areas, each with representative earlier papers in footnotes 2 to 6: scalable oversight; adversarial robustness and AI control; model organisms of misalignment; model internals and interpretability; and AI welfare. The opening paragraph names a slightly different set ("Adversarial Robustness, Dangerous Capability Evaluations, and Scalable Oversight"); dangerous-capability evaluations do not appear as a separate area in the list.
What came out of the program
Many later notes in this collection credit the program in their bylines or acknowledgments, among them Subliminal Learning (July 2025), Persona Vectors (August 2025), Open-sourcing circuit tracing tools (May 2025), Petri (October 2025), Inoculation Prompting (October 2025), AuditBench (March 2026) and Model Spec Midtraining (May 2026). The later announcement gives the program's own counts of participants and publications. Which of them came from the first cohort is not stated anywhere read for this note.
Limitations and open questions
The post states intentions and terms; it reports no outcomes. It does not say how many applications were expected, how mentors and projects would be matched, who owns the research outputs, or how the program would be evaluated. The terms are those of the first cohort; the later announcement and job description changed the length, stipend and compute budget (see Anthropic Fellows Program for AI safety research: applications open for our next cohorts).
Research questions
- How many of the first cohort met the stated aim of a (co-)first-authored paper? A list of first-cohort fellows matched against their publications would answer it.
- How did the restriction to public APIs and open-source models shape the research? Counting which models later Fellows papers study, against Anthropic staff papers from the same period, would show the effect of the rule.
- Which of the five listed areas produced the most published work? Classifying Fellows-credited papers by area would test whether the output followed the stated areas.
- How far did the program draw the "mid-career technical professionals" it was aimed at? Published background data on fellows, if released, would show this.
How it connects
- Anthropic Fellows Program for AI safety research: applications open for our next cohorts (December 2025) is the program's later announcement, which this post's July 2026 notice points to.
- Introducing Anthropic's Safeguards Research Team (February 2025) and Anthropic Residency (October 2026) link the program: the first says the team works closely with it, and the second says the Residency builds on it.
- Footnotes 2 to 5 cite papers that have notes here: Measuring Progress on Scalable Oversight for Large Language Models, Many-shot Jailbreaking, Sabotage Evaluations for Frontier Models, Sleeper Agents, Sycophancy to Subterfuge, Simple probes can catch sleeper agents and Towards Monosemanticity.
References
- Anthropic (2024). Introducing the Anthropic Fellows Program for AI Safety Research. Alignment Science Blog, listed under December 2024 (update notice dated July 2026; retrieved 13 November 2026). https://alignment.anthropic.com/2024/anthropic-fellows-program/ record
- Crosspost: Introducing the Anthropic Fellows Program. LessWrong, posted 30 November 2024. https://www.lesswrong.com/posts/xwebc6CLWuLmahcx6/introducing-the-anthropic-fellows-program