Keep the map of ML usage and safeguards across our products current, and lead the security design of new ML features
Break our safeguards: test the boundaries on jailbreaks, prompt injection, prompt leaks and toxicity, and turn findings into eval cases that run on every model, prompt and safeguard change
Design and build mitigations with ML engineers that hold when detection misses
Define the security metrics that show whether our safeguards work, and report on them
Review changes that could weaken a safeguard
Track AI attack research and turn new techniques into tests and mitigations
