Overview
Microsoft Superintelligence team’s mission is to empower every person and every organization on the planet to achieve more. This role is part of Microsoft AI’s Superintelligence Team, a startup-like team inside Microsoft AI, created to push the boundaries of AI toward Humanist Superintelligence — ultra-capable systems that remain controllable, safety-aligned, and anchored to human values. Our mission is to create AI that amplifies human potential while ensuring humanity remains firmly in control, aiming to deliver breakthroughs that benefit society.
We are hiring a Product Manager to own AI model security — the discipline of making our frontier models resilient against adversarial attack and purpose-built for security practitioners. This role has a dual mandate: (1) harden our models against the full spectrum of LLM security threats — prompt injection, data exfiltration, jailbreaking, training data extraction, zero-day exploit generation, model poisoning, and agentic workflow exploitation — and (2) partner closely with Microsoft Security product teams (Azure Security, Security Copilot) to ensure our models deliver best-in-class capabilities for real-world security workflows.
This is a security role where you will think like an attacker, understand the OWASP LLM Top 10, and apply product judgment to tradeoffs between model capability and attack surface. You will also understand what security analysts and incident responders need from AI, working backwards from their workflows to define model training priorities, evaluation benchmarks, and product requirements. You will work with model researchers, engineers, and red teamers, personally building evaluation frameworks, defining security benchmarks, and driving decisions about what to ship. This is a small team with high ownership, where you will see your work in production and be accountable for outcomes.
Responsibilities
- Own the model security roadmap: Define and prioritize the security hardening strategy for our frontier models across the full OWASP LLM threat surface — prompt injection (direct and indirect), data exfiltration, jailbreak resistance, system prompt leakage, training data extraction, and adversarial manipulation of agentic workflows.
- Drive zero-day and exploit defense: Work with researchers to evaluate and mitigate the risk of models being used to generate zero-day exploits, malware, or novel attack vectors. Define thresholds, build evaluation datasets, and own the decision framework for what the model should and should not be capable of in the security domain.
- Build and scale red-teaming frameworks: Design, run, and iterate adversarial testing programs — both automated and human-driven — to continuously probe model vulnerabilities. Establish metrics (e.g., jailbreak success rate, injection bypass rate, exfiltration resistance) and drive measurable improvement over time.
- Partner with Microsoft Security product teams: Work closely with Azure Security and Security Copilot teams to translate their product requirements into model training priorities. Ensure our models are purpose-built for threat detection, incident triage, vulnerability assessment, log analysis, and compliance reasoning.
- Define security-specific model evaluations: Build benchmark suites and evaluation frameworks that measure real-world security usefulness — not just academic performance. Drive training data strategy to improve domain-specific model quality for security practitioners.
- Shape security policy and launch readiness: Establish clear security criteria for model launches. Own the security dimension of go/no-go decisions, with frameworks that balance capability, risk, and deployment context.
- Stay at the frontier: Track the rapidly evolving LLM security landscape — new attack techniques, emerging standards (OWASP, NIST AI RMF), regulatory requirements (EU AI Act), and academic research. Translate what you learn into actionable product priorities.
- Influence model training and architecture: Partner with researchers and engineers to embed security considerations into model training, fine-tuning, RLHF, and post-training safeguards. You don’t just test — you shape what gets built.
Required Qualifications
- Bachelor’s Degree AND 5+ years experience in product management, security engineering, or software development OR equivalent experience.
- Demonstrated hands-on experience with AI/ML systems — you have personally built, evaluated, or shipped ML-powered products or security tools.
- Deep familiarity with LLM security threats: prompt injection, jailbreaking, data exfiltration, adversarial attacks on generative models — through professional experience, red-teaming, or security research.
- Experience defining product requirements and driving decisions in partnership with researchers or ML engineers.
- Track record of building evaluation systems, security benchmarks, or adversarial testing frameworks — not just consuming them.
- Ability to operate autonomously, make decisions with incomplete information, and drive projects from ambiguity to shipped outcomes.
Preferred Qualifications
- Technical background in computer science, security, or AI/ML — a postgraduate degree is a plus but not required.
- Experience in offensive security, penetration testing, or red teaming — ideally applied to AI/ML systems.
- Familiarity with security workflows and tooling (SIEM, SOAR, EDR, threat intelligence platforms) and how practitioners use them in production.
- Understanding of the model lifecycle (pre-training, fine-tuning, RLHF, deployment, monitoring) and where security interventions are most effective.
- Experience working with or within enterprise security organizations (e.g., Microsoft Security, CrowdStrike, Palo Alto Networks, or similar).
- Published research, blog posts, or public contributions in AI security, adversarial ML, or LLM red teaming.