Design, build and ship safeguards in the agent harness for prompt injection, single and multi-turn jailbreaks, toxicity and sensitive topics, and sensitive data exposure
Build evals that measure safeguard efficacy on both false and true positives, using datasets from customer reports, adversarial testing and production traffic
Work closely with our Security team: turn their adversarial findings into eval cases and fixes, and build mitigations together that hold when detection misses
Select, integrate and fine-tune detection models within a stack shared by our commercial and government environments
Balance quality, latency and cost for controls that run in the request path
Build configurable safeguard policies so customers can decide how sensitive topics are handled
Monitor safeguard behavior in production and fix regressions
