AssuranceRemoteRef: DL-2026-004
Red-Team Engineer, Adversarial Robustness
Design attack suites for injection, escalation and goal hijacking against systems we are about to ship, and track regressions release to release so a closed hole stays closed.
What you'd do
- Design and run attack suites against agentic systems: prompt injection from untrusted content, privilege escalation through a tool an agent was given more access to than it needed, goal hijacking, and plain tool misuse.
- Build regression tracking so a fix that closes one attack path does not quietly reopen an adjacent one on the next release, the way a patched extension permission model can still be walked around through a sibling API.
- Report findings the way a client's own security team expects: reproducible steps, a severity that reflects real impact rather than theoretical worst case, and a concrete mitigation, not just a writeup of the exploit.
- Stay ahead of published attack techniques against agentic systems and adapt them to the specific tools, permissions and trust boundaries of what we are actually shipping.
- Work adversarially against your own team's systems before a client, or someone with worse intentions, does it for you.
What we're looking for
- Applied experience red-teaming LLM or agentic systems specifically, or a strong adjacent background in application security with a demonstrated ability to pick up agent-specific attack surface fast.
- Can write an exploit, and can also write it up so an engineer who was not in the room can reproduce it, understand why it worked, and fix it without you.
- Comfortable working adversarially against your own team's systems: the goal is to find the hole before a client does, not to prove a point.
- Familiarity with common agent frameworks, their tool-calling conventions, and where their trust boundaries actually sit versus where the documentation claims they sit.
- Judgment about severity: not every prompt injection is an incident, and treating all of them as one buries the ones that are.
Ready to apply?
Fill in the form on the next page, CV included; one submission sends it straight to us.