Conduct rapid, security assessments on agentic services, focusing on vulnerabilities unique to Generative AI such as prompt injection, tool-use escalation, and autonomous lateral movement. Engineer and execute attack sequences that exploit non-deterministic model behaviors, agentic logic errors, and data poisoning vectors. Write code to transform manual vulnerability discoveries into automated regression testing frameworks ("Auto Red Teaming") that prevent regression in future model versions.
Partner directly with developers during the design and build phases to provide immediate feedback, effectively shortening the feedback loop between offensive findings and defensive engineering. Maintain and expand a library of agent-specific attack patterns and exploit primitives to establish release criteria for new models.
Job description provided by Google.
View the original posting
