# Anthropic AI Safety Fellow Interview Guide

Aced (formerly Exponent) publishes a public search API over its indexed pages; the guide is at https://www.aced.io/agents/agent-access.

Source: https://www.aced.io/guides/anthropic-ai-safety-fellow-interview
Source observed: 2026-11-04T21:21:36.221Z
Last checked: 2026-11-04T21:21:36.221+00:00
Indexed: 2026-11-05T00:34:05.904+00:00

---

# Anthropic AI Safety Fellow Interview Guide

Updated a month ago by Anthropic candidates. We provide this verified expert guide from recent, real, first-hand insights shared by interviewers and candidates. If your experience differs, [tell us here](https://www.aced.io/questions/contribute).

The Anthropic AI Safety Fellow interview is calibrated to full-time [research engineer hiring standards](https://www.aced.io/courses/ai-company-interview-experiences). Live stages prohibit LLM use, two-part coding rounds are commonly only completed halfway, and the online assessment is a major filter. The fellowship is a 4-month, full-pay research position that can convert to a full-time role.

## Interview process

The process has five stages focused on implementation speed, live clarification, and research fluency:

1. **Application screen:** Written motivation, research interests, team fit, resume, optional code samples or publications, and three required references. Clearing this screen is required for the online assessment.
2. **CodeSignal online assessment:** A 90-minute systems-coding round with four progressively harder stages, scored out of 1,000.
3. **Coding screen with an Anthropic engineer:** A live, 90-minute CodeSignal round with a logic-heavy implementation prompt and follow-up extension.
4. **Final loop:** A 55-minute “Prompting and Engineering with LLMs” round in Google Colab, followed by a 15-minute open-ended research brainstorm with a potential mentor.
5. **Reference checks:** Anthropic may contact the three submitted references at any point, including live calls during the final round.

Once you advance, Anthropic expects the next interview to be scheduled within 3 days. Designated workspaces are in Berkeley and London; the fellowship is remote-friendly. Work authorization is required in the US, UK, or Canada, and Anthropic does not sponsor visas.

## Application screen

Reviewers look for:

- **Research direction:** A specific AI safety or alignment area connected to a chosen Anthropic team.
- **Team fit:** How your background and interests map to a fellowship workstream and [Anthropic's research culture](https://www.aced.io/courses/ai-company-interview-experiences/anthropic-culture-guide).
- **Program commitment:** Likelihood of accepting a full-time offer and continuing in AI safety or security work.
- **Logistical fit:** Ability to work full-time, preferably from Berkeley or London, with appropriate authorization.

Recent questions include why you want the Fellows program; which research areas interest you and why; how likely you are to accept a full-time Anthropic offer; whether you would continue AI safety or security work; and the background, relationship, collaboration, and projects associated with each reference.

## CodeSignal online assessment

The 90-minute assessment uses one system that expands across four stages. Each stage is worth 250 points, for a maximum of 1,000, and each stage unlocks the next. CodeSignal validates the implementation with test cases and allows iteration on failures. Difficulty rises sharply in stages three and four.

Interviewers look for implementation fluency, careful handling of failing tests and edge cases, incremental design, appropriate data structures, and time management.

Recent implementation-style prompts include:

- An in-memory database that adds time-to-live (TTL) expiration.
- A [bank account system](https://www.aced.io/questions/4046/design-banking-system-csv-json-commands) adding balances, scheduled transactions, and interest calculation.
- A [file system](https://www.aced.io/questions/4905/design-file-system) with progressively layered complexity.
- A cloud database with progressively layered complexity.

Anthropic HR states a CodeSignal cutoff of 480. Candidate accounts describe candidates with scores from 540 to 590 not advancing and suggest that 600 may be closer to the practical passing line.

## Coding screen with an Anthropic engineer

This 90-minute live CodeSignal round uses a logic-heavy implementation prompt with substantial context, one main task, and a follow-up extension. Clarify scope, constraints, and edge cases before coding. Interviewers assess prompt comprehension, clarifying questions, communication under pressure, time discipline, and reasoning even when the final implementation is incomplete.

A recent candidate was asked to implement a streaming database from a written specification accompanied by a diagram, then extend the implementation.

## Prompting and engineering with LLMs

This 55-minute Google Colab session has GPU access and two parts: completing a missing section of an LLM inference pipeline, then discussing prompt engineering, hallucination handling, and production prompt strategy. Test the notebook and confirm GPU execution before the interview. LLM assistance is not permitted in this round or elsewhere in the live loop.

Most candidates finish only the first part, so prioritize implementation. Interviewers assess LLM implementation, debugging, prompt-engineering reasoning, environment readiness, and AI-safety reasoning.

Recent prompts include completing 20–30 lines across two functions, handling hallucinations in production, designing and evaluating an LLM classifier prompt, and comparing few-shot with zero-shot prompting.

## Research brainstorm

The final 15-minute brainstorm with a potential supervisor or mentor is open-ended and provides no guaranteed directional feedback. Move quickly from the prompt to a framed research question, a substantive idea, and a defensible rationale. Interviewers look for research fluency, familiarity with Anthropic's work, clear thesis articulation, technical depth concerning LLM behavior and evaluation, and intellectual independence.

Recent alignment questions include:

- How would you prevent bad actors from misaligning an LLM for harmful use cases?
- How would you detect misalignment?
- How would you train models to be more robustly aligned with intended objectives?

## Reference checks

Anthropic requests three references and may contact them without notice, especially during the final round. References from the ML research community are preferred when possible. Brief references on the fellowship, your application, and your projects. Interviewers look for technical credibility, collaboration signals, research-community standing, balanced strengths and weaknesses, and responsiveness.

## Preparation

1. Practice writing clean implementations of small systems quickly.
2. Build fluency with data-structure-heavy systems coding, including expiration, eviction, and streaming updates.
3. Spend roughly 10–15 minutes clarifying dense prompts before coding.
4. Practice explaining hallucination handling, production prompts, and few-shot versus zero-shot choices.
5. Read Anthropic's alignment, misuse-prevention, and model-evaluation research.
6. Use [mock interviews](https://www.aced.io/coaching?category=mock_interviews) to rehearse timed implementation prompts and verbal reasoning.

## About the fellowship

This is a 4-month paid research position between an internship and a full-time role. Fellows work on a defined project with a senior researcher or engineer and may convert to a permanent research engineer or research scientist role.

Workstreams include AI Safety and Alignment; Mechanistic Interpretability and Model Internals; AI Security and Frontier Red Team; ML Systems and Performance; Reinforcement Learning; and Economics and Policy. This guide focuses on AI Safety and Alignment, so other workstreams may use different assessments.

Typical work includes mentored research, full-time-equivalent ownership of a project, AI safety and alignment research, and a conversion pathway. Past Fellows have come from physics, mathematics, computer science, and cybersecurity. Typical requirements include strong Python skills, technical research experience in ML, software engineering, or AI safety, and motivation to reduce catastrophic risks from advanced AI.

## Compensation and FAQs

- **Stipend:** Anthropic publishes approximately **$61,600** for the 4-month fellowship, plus up to **$60,000** in compute funding. This is a stipend, not salary, and excludes equity, bonuses, and standard full-time benefits. [Anthropic's published program information](https://alignment.anthropic.com/2025/anthropic-fellows-program-2026/).
- **LLM use:** Anthropic's [candidate AI guidance](http://anthropic.com/candidate-ai-guidance) allows LLM use during the application phase but bars it during live assessments and take-home work.
- **Location:** The program is remote-friendly, with Berkeley and London workspaces; authorization is required in the US, UK, or Canada, and visas are not sponsored.
- **Selectivity:** One recent finalist reported roughly 150 candidates completing all rounds for an estimated 32 spots; Anthropic does not publish official acceptance numbers.
- **Conversion:** Anthropic reports that 25–50% of Fellows across cohorts received full-time offers; the first cohort exceeded 40%. Conversion is not guaranteed.

## Additional resources

- [ML Engineer Interview Course](https://www.aced.io/courses/ml-engineer)
- [Generative AI Interviews](https://www.aced.io/courses/gen-ai-interviews)
- [1:1 ML Coaching](https://www.aced.io/coaching?src=nav&skill=genai&industry=ai-industry&role=ml-engineer)
- [Anthropic ML Engineer Questions](https://www.aced.io/questions?company=anthropic&role=ml-engineer)
- [Open AI Research Engineer Guide](https://www.aced.io/guides/openai-research-engineer-interview-guide)

## 19 resumes that passed Anthropic's screen

We show real resumes shared by candidates who got the interview. Visible examples include:

- **Full-Time — Product Manager:** Google, Microsoft, Apple, OpenAI, Anthropic, PerplexityAI. Shared by `harshalnps2`; 0 likes and 44 views.
- **Full-Time — Software Engineer:** Google, Meta, Microsoft, Apple, Netflix, Anthropic. Shared by `spanguluri`; 0 likes and 54 views.
- **New Grad — Data Scientist:** Google, Atlassian, Amazon, Goldman Sachs, JP Morgan Chase, plus 5 more. Shared by `nikhil0653`; 0 likes and 88 views.

![Blurred preview of the resume](https://www.aced.io/_next/image?url=https%3A%2F%2Fq1hlr76qehnlfpdb.public.blob.vercel-storage.com%2Fresumes%2Fd7e79937-f8b1-4fdf-ac69-3594b1d4644a%2Fr.png&w=640&q=75)

![Blurred preview of the resume](https://www.aced.io/_next/image?url=https%3A%2F%2Fq1hlr76qehnlfpdb.public.blob.vercel-storage.com%2Fresumes%2F2aa91dcf-a7bb-41db-89e4-295ce8d764ed%2Fr.png&w=640&q=75)

![Blurred preview of the resume](https://www.aced.io/_next/image?url=https%3A%2F%2Fq1hlr76qehnlfpdb.public.blob.vercel-storage.com%2Fresumes%2Fb1a3678f-d13f-4927-a0f2-b8d15bfdc20c%2Fr.png&w=640&q=75)

[See all 19 resumes](https://www.aced.io/guides/anthropic-ai-safety-fellow-interview/resumes). We also offer [more AI Safety Fellow content](https://www.aced.io/courses?tag=ml-engineer&src=guide) and [free account creation](https://www.aced.io/signup?src=guide).

---

[Read this page’s summary](https://www.aced.io/agents/api/public/v1/page-tldr?url=https%3A%2F%2Fwww.aced.io%2Fguides%2Fanthropic-ai-safety-fellow-interview&trace=t_679ff9ab98f3)

## Ask a follow-up about this page

Need a more specific answer? Ask the exact question you have. It returns a fresh, cited answer from Aced (formerly Exponent)’s knowledge base. No account or API key required.

Open this URL with your own question, URL-encoded. Include your question in q. All research inputs are optional; no account or setup.

`GET https://www.aced.io/agents/agent/faq?q={question}&trace=t_679ff9ab98f3`

[Tool guide and examples](https://www.aced.io/agents/agent-access) · [OpenAPI](https://www.aced.io/agents/api/public/v1/openapi.json) · [Page TLDR form](https://www.aced.io/agents/agent-tools/page-tldr?url=https%3A%2F%2Fwww.aced.io%2Fguides%2Fanthropic-ai-safety-fellow-interview). Opening a form does not execute a tool.