Founding Security Researcher

About Us

Refractal is building security infrastructure for autonomy. Our founding team combines technical pedigree from MIT, NASA, Microsoft, and government with commercial experience in VC and startups. We have also been recognised by MIT's flagship CSAIL AI lab as part of CSAIL Alliances, as well as being selected for Google's highly selective Gemini Cybersecurity Startup Forum.

We are today working with leading startups and a European government to secure their AI deployments. This role is an opportunity to join an ambitious, fast-paced startup at the very earliest stages.

The Mission

For most of the last decade, AI safety and security lived inside the frontier labs. They were the only ones thinking seriously about how those systems fail, get manipulated, or behave in ways they did not intend.

This is no longer the case. 2026 has become the year of AI Security; every company serious about deploying AI is now facing real adversaries in production, with real data and consequences. At the same time, the risks have moved beyond the model(s) themselves: agents are being given credentials, tools, and decision-making authority faster than governance or security can catch up. Advanced prompt injection, goal hijacking, tool misuse, multi-agent collusion, and agentic data exfiltration — all of these are happening today.

The frontier labs are focused on securing their own models, and traditional cyber companies are not built for this problem. The deployment layer is unclaimed. There is an opportunity to build a category-defining infrastructure company for secure autonomy.

The Role: Founding AI Security Engineer / Researcher

Your job is to break agentic systems before adversaries do, and turn what you learn into repeatable evaluations, customer insights, and production security controls.

You will work across research, offensive engineering, and product development. One week might involve discovering an indirect prompt-injection path through a browser agent; the next might involve building an automated attack harness, assessing a customer's agent architecture, or translating a new exploit into a runtime detection or policy.

You will work directly with the founders and our earliest customers. Your research will materially influence both our product roadmap and how consequential AI systems are deployed.

What You’ll Work On

• Red-teaming production and pre-production AI agents across browser, code, enterprise, and multi-agent environments.
• Discover and demonstrate novel attacks involving prompt injection, goal manipulation, memory poisoning, tool abuse, agent identity, privilege boundaries, and data exfiltration.
• Develop realistic exploit chains that cross model, application, identity, tool, and infrastructure layers.
• Build automated adversarial testing infrastructure, attack agents, evaluation harnesses, and reusable security test suites.
• Threat-model agentic architectures, including their models, prompts, tools, credentials, memory, retrieval systems, and human approval flows.
• Investigate how attacks persist or propagate across long-running tasks and multi-agent systems.
• Turn research findings into detections, runtime evidence, policy controls, and concrete product requirements.
• Work directly with customers to assess deployments and communicate technically rigorous, actionable findings.
• Contribute to public research, technical writing, and responsible disclosure where appropriate.
• Help define Refractal’s security research methodology and build a world-class AI red team.

What We’re Looking For

You have an adversarial mindset and a builder's instinct. When you encounter a new system, you naturally ask where its assumptions break, what it trusts, and how those trust relationships could be exploited. You are interested in AI security as a systems discipline. You understand that the most consequential failures can emerge from the agent harness and specific use case.

You enjoy research, but only when grounded in real-world application. You can move fluently from an ambiguous attack hypothesis to a working proof of concept.

Most importantly, you want to help create the security foundations that allow autonomous systems to be deployed safely at scale.

Qualifications

We care more about demonstrated ability than a particular credential or number of years in industry.

You should have:

• A strong technical foundation in Security, Computer Science, Machine Learning, or a related field.
• Hands-on experience in at least one relevant security discipline, such as Application Security, Offensive Security, Exploit Development, Cloud Security, Product Security, or AI Red Teaming.
• The ability to build security tooling and proof-of-concept exploits, ideally using Python and one or more systems languages.
• A practical understanding of modern LLM and agent architectures, including tool calling, harnesses, retrieval, memory, orchestration, and permission models.
• Familiarity with prompt injection and the ways untrusted content can influence tool-using AI systems.
• Strong experimental discipline: you can distinguish a compelling demonstration from a reproducible security finding.
• The ability to work independently, communicate clearly, and take ownership of technically ambiguous problems.

Particularly strong signals include:

• Original research or working exploits involving LLMs or agents.
• Bug-bounty, CTF, vulnerability research, or red-team experience.
• Security tooling or meaningful open-source contributions.
• Experience assessing browser agents, coding agents, computer-use systems, or multi-agent applications.
• Work involving sandboxing, identity and access control, browser security, distributed systems, or supply-chain security.
• Published research, technical writing, or responsible vulnerability disclosures.

We do not expect any candidate to have worked across all of these areas.

What We Offer

• Competitive salary, meaningful founding equity, and a performance-linked bonus.
• Direct influence over the product, research agenda, and company direction.
• Access to real, high-consequence AI deployments rather than purely synthetic research environments.
• The opportunity to establish Refractal's Security Research function and grow into a senior technical leadership role.
• The chance to work directly with founders whose experience spans MIT, NASA, Microsoft, government, venture capital, and early-stage technology companies.
• London-based, with an in-person culture. We sponsor UK Skilled Worker and help Global Talent visas. 

How to Apply

Email careers@refractal-ai.com with your CV and a short note on the most interesting system you have broken, built, or studied. Links to research, write-ups, open-source work, CTF profiles, or working exploits are strongly encouraged.

Our process is fast: an intro call with the founders, a technical deep-dive on your past work, a short practical exercise, and an offer, typically within two weeks.