Arena is the platform for evaluating how AI models perform in the real world. Founded by researchers from UC Berkeley's SkyLab, we're on a mission to measure and advance the frontier of AI for real-world use, and to build the foundation for everyone to understand, shape, and benefit from it.
Tens of millions of people use Arena each month to evaluate how frontier systems handle the work they actually do. The preferences they share power the most transparent, rigorous, and human-centered evaluations in AI. Leading AI labs, enterprises, and independent researchers rely on our work and open datasets to understand how models behave in real workflows: agentic coding, creative generation, professional productivity, and beyond. We go beyond leaderboards and decompose what human experience reveals about AI, so models advance toward the work people actually do.
We're a team of researchers, academics, builders, and creatives from UC Berkeley, Google, Stanford, and DeepMind. We seek truth, move fast, and value craftsmanship, curiosity, and impact over hierarchy. We're building a company where thoughtful, curious people from all backgrounds can do their best work together, in an office culture that radiates excellence, energy, and focus.
This is an engineering role. The work is in the product code: the authorization and identity layer, the API surfaces our users and partners depend on, and the libraries and patterns other engineers build on top of. You will design systems, write production code, and ship on the same cadence as the rest of engineering.
Tens of millions of people use Arena every month to compare frontier models side by side. That makes our product surface an unusual one to defend: consumer scale, high-value partner data, frontier model integrations, and a user population that probes it constantly.
You would define how product security works here, including the patterns, primitives, and defaults other engineers inherit.
You'll work closely with product, infrastructure, and platform engineers. The work is zero-to-one in places and scale-it-up in others. We move fast and stay rigorous.
What You'll DoOwn authorization and identity in the product. Design and ship the authentication, authorization, and multi-tenant isolation code across our platform, along with the account, session, device, and credential lifecycle that other systems are built on: issuance, binding, expiry, and revocation. This is code you write, not findings you route.
Solve AI security problems the industry has not settled. Untrusted model output crossing trust boundaries, tool-use and agentic surfaces including sandboxed code execution, isolation between model providers, and the authorization and handling controls around evaluation data. There is little prior art here and no off-the-shelf playbook. You will help write ours.
Build privacy-preserving systems, and implement the data lifecycle in code. Pseudonymous identifiers, secret-backed derivation, and controlled re-identification, with the judgment to tell real pseudonymization from the appearance of it. Retention, deletion, and minimization as working product behavior across every store and pipeline that holds user data, including proving that what should be gone is actually gone.
Make the product explain itself after the fact. Security-relevant events, including authentication, account changes, and privileged actions, need to be emitted from product code as durable, queryable records. You design what gets written, and you write it.
Ship security-sensitive changes end to end. Request paths, partner integrations, data models, telemetry pipelines, and rollout controls that let you turn a change off without a deploy. That includes production migrations and backfills for sensitive data, throttled, resumable, observable, and reversible. It also includes the product integrations for third-party security and trust services, where you design the call paths and define conservative failure behavior.
Own application security across the product, its APIs, and the browser. Content Security Policy and its reporting path, security headers, cookie and origin policy, and the third-party JavaScript running alongside user content. Ship the libraries, middleware, tests, and platform patterns engineers adopt because they are the easiest path. Lead threat modeling early enough to change a design, and own the fix in the codebase rather than the ticket.
What We're Looking ForWe care most about two things: recent hands-on production engineering, and code-level application security judgment.
Production code you have personally written recently that you can walk through in depth: the architecture, how it changed over time, and which parts were yours. We hire software engineers for this role, and the engineering bar is the same as the rest of the team.
Application and API security: you apply it at the code level, not the checklist level. Authorization models, session and token handling, and multi-tenant isolation, with a specific problem you found or fixed that you can explain down to the mechanism.
6+ years of software or security engineering experience, with meaningful time building and shipping production systems at scale.
Strong proficiency in a modern backend language, and the judgment to design interfaces other engineers will live with for years.
Solid data fundamentals. You're comfortable modeling and querying Postgres, you know where security decisions belong in a data layer, and you have run migrations against production data without breaking it.
Working fluency with cryptographic application primitives: HMACs, secure random generation, key and secret rotation, and the ways sensitive data leaks through logs, traces, and stored payloads.
Exposure to securing AI, agentic, or LLM-backed systems in production, or the appetite to go deep on it quickly. This is where the newest and least-solved work sits.
Multi-tenant systems experience where isolation is a hard requirement, and the tenants have competing interests.
Offensive instincts. You can find the flaw yourself, and you would rather build the thing that finds all of them.
Threat modeling and secure architecture design on systems you did not build. You can read an unfamiliar codebase, identify the security-critical design assumptions, and change the design before it ships.
A product-oriented mindset. You think about the developer experience of what you build, and about what a security control costs the people who use it. You ask "why" before "how."
The ability to turn legal and policy requirements into working systems. Retention, deletion, user rights, and partner obligations arrive as constraints, and you translate them into data models and product behavior that hold up under scrutiny.
Clear communication. You routinely build alignment across engineering, product, legal, and leadership.
Comfort with ambiguity. We're a startup. Scope is fluid, context shifts, and you'll wear many hats.
TypeScript
Go
Python
Postgres
Hono
Next.js + React
Terraform + Spacelift
Vitest
We offer competitive compensation and equity aligned to the markets where our team members are based. The base salary range will depend on the candidate’s permanent work location.
Comprehensive health and wellness benefits, including medical, dental, vision, and additional support programs.
The opportunity to work on cutting-edge AI with a small, mission-driven team
A culture that values transparency, trust, and community impact
Come help build the space where anyone can explore and help shape the future of AI.
Arena Intelligence provides equal employment opportunities (EEO) to all employees and applicants for employment without regard to race, color, religion, sex, national origin, age, disability, genetics, sexual orientation, gender identity, or gender expression. We are committed to a diverse and inclusive workforce and welcome people from all backgrounds, experiences, perspectives, and abilities.
.png){kind=logo}
