Anthropic Fellows Program 2027 - AI Safety Research

A funded research programme that pays technical people to work full-time for four months on an empirical AI research project with mentorship from an Anthropic researcher. Fellows use external infrastructure such as open-source models and public APIs, and the aim is a public output, usually a paper submission. It began as a pilot announced in December 2024 for a cohort of 10 to 15 fellows working on AI safety, and has since expanded into five workstreams covering safety, security, ML systems and performance, reinforcement learning, and economics and policy.

The stipend is 3,850 USD, 2,310 GBP or 4,300 CAD per week, depending on where you are based, with benefits that vary by country. On top of that, fellows get funding for compute and other research expenses at roughly 15,000 dollars a month. The expectation is 40 hours a week across four months, with a possible extension. The pilot cohort paid 2,100 USD a week with about 10,000 dollars a month of research budget, so any figure you see in that range is from the older announcement rather than the current one.

Only if they already hold work authorisation in the US, the UK or Canada and will be located in that country during the programme. Anthropic states clearly that it is not currently able to sponsor visas for fellows, and that the visa sponsorship policy on its full-time job postings does not apply to the Fellows Program. If you would need sponsorship to work in one of those three countries, you are not eligible, and no part of the application process changes that.

No. Anthropic says directly that you do not need a PhD, prior machine learning experience or published papers, and that successful fellows have come from physics, mathematics, computer science, cybersecurity and other quantitative backgrounds. The programme was built for people moving into empirical AI research rather than for people already established in it. The published requirements are fluent Python, availability to work full-time, and a strong technical background.

There is no published deadline. Anthropic runs multiple cohorts a year and reviews applications on a rolling basis, and the open postings say they are accepting applications for a cohort expected to start in January 2027. Late November 2026 is a reasonable planning target for that intake rather than an announced cutoff. An earlier blog announcement listed a 26 July close for a November 2026 cohort, which has passed, so treat dates in that post as historical.

No. Fellows are not officially employed by Anthropic, and the postings state that Anthropic's policies for full-time roles, including the office attendance expectation and visa sponsorship, do not apply to the programme. It is a fixed-term funded research fellowship with a stipend, a compute budget and a mentor. Anthropic also says it does not guarantee that any full-time offers will follow, although in previous cohorts 25 to 50 percent of fellows received one.

Empirical projects aligned with Anthropic's research priorities, chosen with a mentor who pitches ideas. Past fellows worked on agentic misalignment, stress-testing 16 frontier models in simulated corporate settings; on subliminal learning, where models pass behavioural traits through unrelated data; on rapid response methods that block whole classes of jailbreaks after a handful of examples; on open-source circuit tracing for interpretability; and on security agents that identified 4.6 million dollars of blockchain smart contract vulnerabilities.

Anthropic does not publish an acceptance rate or a current cohort size, so anyone quoting one is guessing. What it does publish is the outcome side: in one earlier cohort more than 80 percent of fellows produced papers, over 40 percent of the first cohort went on to join Anthropic full-time, and across previous cohorts 25 to 50 percent of fellows received a full-time offer. The selection runs through a reference check, technical assessments and interviews, and a research discussion.