Anthropic Fellows Program for AI safety research: applications open for our next cohorts

The Anthropic Fellows Program's later announcement on the Alignment Science Blog (December 2025, revised since), which reports that more than 200 fellows have taken part since early 2025, that over 65% have published work from the fellowship and over 35% later joined Anthropic, describes four-month projects with about $15,000 a month of compute, and gives examples of Fellows work in security, interpretability and model organisms.

Anthropic (Alignment Science Blog) · 2025-12-21 · original

The Anthropic Fellows Program pays people outside Anthropic to do AI safety research with Anthropic mentors. It began as a pilot announced in late 2024 (Introducing the Anthropic Fellows Program for AI Safety Research). This Alignment Science Blog post is the program's later announcement: it reports how the program has grown, describes how projects work, and opens applications for further cohorts. The blog's index lists it under December 2025 with the label "Anthropic Fellows Program 2026"; the first Wayback Machine capture is from 21 December 2025, so this note gives that as the latest possible publication date. The page has been revised since: the Wayback Machine holds captures with several different contents between December 2025 and October 2026, and the text read for this note (retrieved 13 November 2026) asks for applications by 18 October for a January cohort. The earlier versions could not be retrieved for this note, so the numbers below are those of the current text, which gives no date for them.

What the post reports about the program

The post's figures, all from its opening section and "Program information", are the program's own counts, with no list of fellows or method given:

  • "more than 200 fellows" have taken part "since early 2025";
  • "Over 65% of fellows have published work from their fellowship";
  • "Over 35% of fellows from past cohorts subsequently joined Anthropic to work full-time on AI safety".

The post adds that the program has widened its mentors' areas to "scalable oversight, adversarial robustness and AI control, model organisms, mechanistic interpretability, AI security, and model welfare". Against the first announcement, AI security is new to the list.

Terms as stated

  • Length. Four months of full-time research (the pilot was six).
  • Projects. Mentors "pitch" project ideas; fellows choose and shape a project with the aim of a public output such as a paper.
  • Compute. About 15,000 USD a month (the pilot gave about $10,000 per fellow per month for all research costs).
  • Stipend. Not stated in the post, which refers readers to the job description.
  • Eligibility. The post says it cares "much more about your ability to execute on research than your credentials", and that applicants do not need a PhD, prior ML experience or papers.

The linked job description is a separate, living page. As retrieved on 13 November 2026 it gives a weekly stipend of 4,615 USD (or 3,440 GBP, 6,470 CAD, 4,110 EUR), compute of about $15,000 a month, two workstreams (AI Safety Fellows and AI Security Fellows) with named mentors, work authorization in the US, UK or Canada with no visa sponsorship, and applications by 30 October for a January cohort. It names Constellation as the recruiting and program partner. It also reports two figures measured differently from the post's: "over 80% of fellows produced papers" in "one of our earlier cohorts", and "25-50% of fellows received a full-time offer" in previous cohorts. The post's 35% counts fellows who joined; the job description's range counts offers.

Examples of Fellows work the post gives

The post's section "What fellows work on" attributes these results to fellows; each has a note here, which gives the result's own conditions:

  • Security. Agents that, in the post's words, "identified 4.6M USD in blockchain smart contract vulnerabilities" and found two new zero-day vulnerabilities (AI agents find $4.6M in blockchain smart contract exploits). The collection's note on that study records what the figure measures: simulated exploit revenue, summed over the best model on each of 19 contracts that had already been exploited in real attacks after the models' knowledge cutoffs. The two zero-day vulnerabilities were worth $3,694 in simulated revenue. The post also credits "an Anthropic fellow" with Rapid Response and calls it "a key component of Anthropic's ASL3 deployment safeguards". That preprint is dated November 2024, before the first cohort's onboarding in March 2025; the post does not say under what arrangement the work was done.
  • Interpretability. The open-source release of attribution-graph tools (Open-sourcing circuit tracing tools). The post says fellows "introduced a new method"; the collection's note records that the method was first described by Anthropic in March 2025 and that fellows built the open-source library.
  • Model organisms. Agentic Misalignment, which stress-tested 16 models in simulated corporate settings, and Subliminal Learning. The Agentic Misalignment post (June 2025) describes the work as a collaboration between Anthropic, Aengus Lynch (University College London), Caleb Larson (MATS) and Sören Mindermann (Mila), and its arXiv version (2510.05179v2) lists the same affiliations; neither mentions the Fellows Program, so the attribution to fellows is this post's.

The post's opening also names TASTE, SLEIGHT-Bench, Petri and Model Spec Midtraining as published Fellows work. Because the page has been revised, some of these were added after December 2025 (TASTE, SLEIGHT-Bench and Model Spec Midtraining are from 2026).

Limitations and open questions

The counts are self-reported, with no definition of "published work", no date and no breakdown by cohort. The examples are a recruiting page's selected successes. The page is revised in place without a changelog, so a citation needs the retrieval date.

Research questions

  1. What did the December 2025 version of the page say? Recovering the earliest capture and comparing it with later ones would date each figure and example.
  2. What counts as "published work" in the 65% figure? A list of outputs by type would show how many are peer-reviewed papers, preprints or blog posts.
  3. Did the move from six months to four change output? Publication rates by cohort, before and after the change, would answer it.
  4. How much of the collection's 2025 and 2026 alignment research is Fellows work? Counting Fellows-credited notes against all alignment notes by quarter would show the program's share.

How it connects

References