Refractal is an AI security company building the infrastructure organisations need to defend against autonomous AI threats. 2026 has shown that generally capable intelligence is here. The challenge now is modernising security to stop AI-enabled threats, whether they come from internal agents going rogue or from attackers using AI to become more dangerous.
Refractal was founded on cybersecurity and AI expertise from MIT, Microsoft, NASA and the U.S. Navy. Our product lets organisations detect AI-enabled threats at low cost without sacrificing detection quality. We take a practical approach to AI risk, combining proven cybersecurity methods with frontier AI security research.
Today, we work with governments and organisations that can't afford to get security wrong. We're backed by leading deep-tech and cybersecurity investors and are growing the team to deliver on this mission.
The roleYou'll own the machine-learning models behind Refractal's detections in production, which decide in milliseconds whether an action is safe, and the analysis that shows how well they perform for each customer.
What you'll doTrain, tune and ship classifiers for prompt injection, data exfiltration, policy violations and other risky AI activity.
Design low-latency models, including distilled and small models and rules-plus-ML ensembles.
Analyse production telemetry for drift, blind spots, false-positive cost and new attack patterns.
Build evaluations and benchmarks for our detection models, and work with security engineers to turn new attacks into training data.
Turn company policies and regulatory requirements into measurable controls.
Detection models you own are in production, with quality metrics tracked per customer.
You've reduced false positives or latency measurably without losing detection quality.
Production data feeds a regular cycle of model improvements.
5+ years in applied ML or data science, with models shipped to production.
Strong Python, PyTorch or similar, and solid statistics for evaluation.
Hands-on NLP and LLM work: fine-tuning, embeddings, classification and LLM-as-judge.
Experience measuring fuzzy problems, including labelling, agreement and drift.
Clear communication of trade-offs to engineers and customers.
Experience with adversarial ML, safety or content-moderation models.
Published work on evaluations or AI security.
Radical transparency. We believe the best work happens when everyone feels able to give and receive honest, constructive feedback.
Public service. We see AI security as a form of public service. Getting the AI transition right means putting the right infrastructure in place to manage the threats that come with it, and our work helps protect the institutions and services society depends on.
Fail fast. We test ideas early, seek feedback and learn quickly. We experiment rapidly and change course when something isn't working, without compromising on security.
Pragmatism, not dogmatism. We choose approaches based on how well they solve the problem. That means combining established cybersecurity methods with frontier AI research, questioning our assumptions, and changing our minds when the evidence changes.
Compensation and benefitsSalary: £85,000–£145,000
Competitive equity to share in Refractal's growth
Pension contribution
Private medical and dental cover
Visa and relocation support
Skills Required
- 5+ years of experience in applied machine learning or data science
- Experience shipping machine-learning models to production
- Strong Python skills
- Experience with PyTorch or a similar framework
- Solid statistics for model evaluation
- Hands-on NLP and LLM experience
- Experience with fine-tuning, embeddings, classification, and LLM-as-judge
- Experience measuring fuzzy problems, including labeling, agreement, and drift
- Ability to clearly communicate trade-offs to engineers and customers
- Experience with adversarial machine learning, safety, or content-moderation models
- Published work on evaluations or AI security
What We Do
Refractal builds security infrastructure for high-assurance AI, helping governments and enterprises test AI agents before deployment, protect them in production, and meet compliance requirements. Its software monitors agent activity and enforces policies at the point of action, evaluating tool calls and data access and recording tamper-evident audit trails. The platform integrates with existing telemetry and observability tools and helps detect AI-enabled threats across models, tools, and workflows.
