Offensive Security Engineer

About the AI Security Institute

The AI Security Institute is the world's largest and best-funded team dedicated to understanding advanced AI risks and translating that knowledge into action. We’re in the heart of the UK government with direct lines to No. 10 (the Prime Minister's office), and we work with frontier developers and governments globally.

We’re here because governments are critical for advanced AI going well, and UK AISI is uniquely positioned to mobilise them. With our resources, unique agility and international influence, this is the best place to shape both AI development and government action.

The deadline for applying to this role is Sunday 15th November 2026, end of day, anywhere on Earth.

About the team

Security Engineering at the AI Security Institute (AISI) protects our people, partners, models, data and the research infrastructure that AISI's work depends on. We build secure-by-design platforms, automated governance and intelligence-led detection. We work closely with research units and technology teams so researchers can get on with their work.

We're now building an offensive security team and have three roles available.

About the role

We're looking for an experienced offensive security engineer who builds and breaks. You should be able to find and exploit vulnerabilities across cloud, infrastructure and application layers, and want to apply that work to containing and evaluating frontier AI models.

Offensive security testing is one of the key controls behind AISI's model evaluations. You'll run engagements against our evaluation and research infrastructure. That includes the sandboxes, Inspect (our open-source evaluation framework), and the supporting cloud, identity and platform services. Alongside testing, you'll build and own agent-assisted red-teaming harnesses that test this infrastructure continuously.

You'll work directly with the engineers and with research teams on cyber ranges and evaluation tasks, and there's scope to publish and present your work externally.

If you want to attack the systems frontier AI is tested in, use AI agents for offensive work on real targets, and see your findings change how those systems are built, we'd like to hear from you.

What you'll be doing:

• Attacking evaluation sandboxes. Testing the environments that contain frontier models during evaluations, including container and VM escape, egress controls, and attacking on the evaluation host from inside the sandbox.
• Finding vulnerabilities in Inspect. Finding vulnerabilities in our open-source evaluation framework and the infrastructure around it, and coordinating fixes and disclosure with maintainers and downstream users.
• Building and running agent-assisted harnesses. Building, operating and maintaining agent-assisted red-teaming harnesses end to end. Running them, along with manual testing, against every new sandbox and evaluation infrastructure release before evaluations depend on it.
• Triaging and proving findings. Separating real issues from noise in what the agents find and confirming the real ones with working exploits.
• Comparing agents and humans. Running agent and human red-teaming against the same targets, and feeding what you learn back into the tooling.
• Running wider engagements. Planning and running scoped engagements against our cloud, identity, model-access paths and research platform, including joint exercises with our detection team.
• Working with researchers. Helping to build and harden cyber ranges, reviewing evaluation tasks for realism and scoring, and providing human baselines.
• Managing external testing. Scoping and reviewing external penetration tests.
• Working within governance. Keeping model-assisted offensive work within AISI's internal safety governance and approval processes.
• Sharing your work. Where appropriate, publishing research, writing blog posts and giving talks on sandbox security and AI-assisted offensive work.

Key competencies:

• Finding and exploiting vulnerabilities in complex, fast-changing systems
• Building tools that make offensive work repeatable
• Judging which findings are real and worth fixing, including findings from AI agents
• Operating safely and accountably under rules of engagement

Who we're looking for

Essential experience:

You may be a good fit if you have some of the following skills, experience and attitudes:

• Substantial hands-on offensive security experience in red teaming, penetration testing, vulnerability research or exploit development, with findings you can talk through in depth
• Depth in at least two of:
  • cloud and infrastructure (AWS preferred)
  • container, VM or sandbox escape
  • application security
  • vulnerability research and exploit development
• Strong Python and/or another language such as Go, Rust or C, and a track record of writing tools that other people use and that you maintain
• Experience testing live systems safely
• Clear written and verbal communication, including findings that engineers can act on, and the ability to get issues fixed in teams you don't manage
• A serious interest in where frontier AI is heading, and motivation to help make it safe

Desirable experience:

We may also consider experience with: