JD: Design and operate technical guardrails for LLM-based agents including input/output filtering, prompt injection detection, PII redaction, jailbreak resistance, tool-use constraints (restricting which APIs/functions agents can invoke under what conditions), and secret/credential protection to prevent data exfiltration
Build adversarial evaluation harnesses and red-team AI systems to discover failure modes, misuse vectors, and policy violations; create scenario-based test suites covering prompt injection attacks, multi-turn jailbreaks, indirect prompt injection via RAG documents, tool misuse (e.g., calling unauthorized APIs), and social engineering tactics
Translate AI risk and responsible AI policies into engineering requirements and measurable acceptance criteria; establish safety evaluation strategy defining coverage goals (e.g., 95% of known jailbreak patterns blocked), threat models specific to conversational agents (especially tool-calling and memory-enabled agents), and automated safety gates for CI/CD release pipelines
Implement runtime safety monitoring and automated incident response including detection of anomalous agent behavior (excessive tool calls, abnormal session lengths, policy-violating outputs), real-time blocking/throttling, logging for forensic analysis, and alerting workflows integrated with security operations
Conduct full-scope red team assessments simulating adversarial users, create attack paths demonstrating exploitability, and prepare detailed reports with findings, vulnerability classifications, mitigation recommendations, and presentation-ready summaries for leadership and cross-functional stakeholders
Maintain regression suites for known attack patterns (prompt injections, PII leakage, disallowed content generation, tool misuse, data exfiltration), ensure coverage evolves with new threat intelligence, and coordinate with blue team/security operations to improve detection, response capabilities, and defense-in-depth strategies
Partner with compliance, legal, and responsible AI teams to implement safety controls that meet regulatory requirements (financial services regulations), audit logging for model behavior, explainability tooling for safety decisions, and governance-aware delivery practices that balance safety with product usefulness and performance