We’re building an agentic red-team platform that stress-tests guardrail models. Our system spins up swarms of autonomous agents that evolve prompts via RL and genetic algorithms to uncover both bypasses (unsafe → judged safe) and over-blocking (safe → judged unsafe). If you want to ship impactful safety tech fast, this is for you.
The RoleAs our first dedicated ML engineer, you’ll own the design and implementation of the agent framework and its learning loop—taking us from working prototype to a reliable, scalable product. You’ll work across research, engineering, and product to turn ideas into measurable robustness wins for customers.
What you’ll do
Design and build multi-agent architectures for automated red-teaming (task decomposition, memory, coordination, budgeting).
Implement evolutionary search (mutation, crossover, selection) and integrate RL/bandit strategies to optimize a multi-objective fitness (misclassification, novelty, diversity, severity).
Build evaluation loops against guardrail classifiers (e.g., ModernBERT-based) and policy graders; ensure correct labeling and reproducibility.
Develop text-transformation operators (style, translation, obfuscation, role framing, code-block tricks) safely and systematically.
...