Operational Risk

Job Title

Risk, Operational Risk (Artificial Intelligence Coverage), Vice President, Dallas or Salt Lake City  Dallas, Texas, United States

Lead for AI Architecture – Artificial Intelligence Coverage / Operational Risk

Vice President, Dallas/ Salt Lake City

The Operational Risk Department at Goldman Sachs is an independent risk management function responsible for developing and implementing a standardized framework to identify, measure, and monitor operational risk across the firm. The AI Lead for AI Architecture is a specialized role within this framework, dedicated to strengthening the firm's oversight of AI-related risks arising from model development, deployment infrastructure, technical standards, and the internal AI technology stack.

This professional will be responsible for continuously identifying, monitoring, measuring, and assessing operational risks associated with the firm’s AI architecture decisions, including secure‑by‑design principles, model governance within the tech stack, infrastructure resilience, explainability, data quality and drift, prompt injection defenses, and the alignment of technical architecture with the firm’s AI risk appetite. The role ensures that the firm’s AI systems are architected, deployed, and operated in a manner that is secure, resilient, explainable, and compliant with regulatory obligations.

• Identify, monitor, and analyze operational risks arising from the design, development, and deployment of AI systems, with a focus on risks such as inadequate system alignment, lack of explainability, data quality and drift, prompt injection, hallucination and inaccurate outputs, non‑deterministic behavior, bias and discrimination, model overreach/expanded use, reputational risk from AI failures, agent action authorization bypass, tool chain manipulation and injection, agent state persistence poisoning, and multi‑agent trust boundary violations.
  
  Develop evidence‑based challenges focused on improving architectural risk posture.
• Monitor the firm's AI architecture control inventory for sufficiency and completeness, challenging the absence of controls and the implementation of controls within engineering standards. This includes oversight of mitigations such as AI Firewall Implementation and Management, User/App/Model Firewalling/Filtering, AI System Observability, System Acceptance Testing, Data Quality and Classification/Sensitivity, Human Feedback Loop for AI Systems, LLM‑as‑a‑Judge automated evaluation, Providing Citations and Source Traceability, AI Model Version Pinning, Agent Authority Least Privilege Framework, Tool Chain Validation and Sanitization, Agent Decision Audit and Explainability, Multi‑Agent Isolation and Segmentation, Data Filtering From External Knowledge Bases, Preserving Source Data Access Controls in AI Systems, Role‑Based Access Control for AI Data, Encryption of AI Data at Rest, and Quality of Service and DDoS Prevention for AI Systems.
• Champion secure‑by‑design principles across the AI technology stack, ensuring that security, privacy, and risk controls are embedded into AI system architecture from inception rather than retrofitted.
• Conduct data analysis to identify trends and patterns in AI system performance, model behavior, observability telemetry, and security events, augmenting such analysis with qualitative observations to monitor risk‑taking trends through bespoke metrics at firmwide and divisional/sub‑divisional levels. Escalate concerns to senior management when warranted.
• Contribute to divisional and functional risk profile assessments by highlighting AI architecture risk issues and trends to senior divisional managers and the senior Operational Risk management team.
• Conduct evidence‑based scenario analysis by working with stakeholders to develop plausible tail risk scenarios around AI architecture failures, including prompt injection attacks leading to data exfiltration, hallucination‑driven erroneous financial advice, cascading failures in multi‑agent systems, agent authorization bypass leading to unauthorized transactions, data drift causing model degradation, and infrastructure resilience…