Principal AI Security Researcher

Security is one of the most critical priorities for our customers in a world of growing digital threats, regulatory scrutiny, and estate complexity. Microsoft Security aspires to make the world safer by empowering every user, customer, and developer with a security cloud that delivers end-to-end, simplified protection. The Microsoft Security organization advances this mission by helping secure digital technology platforms, devices, and clouds across customers' heterogeneous environments, while also protecting Microsoft's internal estate. Our culture is grounded in a growth mindset, inspiring excellence, and enabling teams and leaders to bring their full potential each day.   The Microsoft Security Research (MSecR) Purple Team sits at the intersection of offense, defense, and intelligence. AI is now everywhere: organizations and people across every industry are adopting large language models, copilots, and autonomous agents, and these systems bring a brand-new attack surface that can be manipulated, misused, and exploited in ways traditional software cannot. In Microsoft Security, we protect AI from these attacks. We are looking for a Principal-level security researcher who combines a real security background with a deep, hands-on understanding of how modern AI works and how it breaks. You will help execute and conduct AI attacks through purple team simulations, using what you find to make AI safer and to inform our security products so they protect customers better.   This role is for someone who wants to shape how AI attacks are simulated and defended at scale. You will set methodology, build tooling, influence product direction, and drive innovation in attacking and defending AI systems (Security for AI).

As a Principal Security Researcher, you will:  Lead the design and execution of purple team simulations aimed at making AI safer, conducting realistic AI attacks whose findings inform our security products and help them protect customers more effectively.  Develop and execute hands-on attacks against AI systems, including:  Direct prompt injection and jailbreaks that bypass system prompts, safety instructions, and alignment.  Indirect (cross-domain) prompt injection (XPIA), where malicious instructions are hidden in untrusted content the model consumes: documents, web pages, emails, tickets, code, and RAG sources.  Multi-turn, crescendo, and social-engineering style attacks that steer a model toward unsafe behavior over a conversation.  Multimodal injection through images, audio, and files that carry hidden instructions.  Guardrail, content filter, and safety-system bypass, including encoding, obfuscation, and adversarial phrasing.  Sensitive data exfiltration and leakage, including system prompt and grounding-data disclosure and cross-tenant or cross-session leakage.  Agentic and tool-use abuse: excessive agency, confused-deputy attacks, unsafe autonomous actions, and manipulation of tools, plugins, and connected systems.  Data and model attacks: training-data and RAG poisoning, model extraction, model inversion, membership inference, and adversarial examples / evasion.  Build scalable AI red teaming tooling and automation (custom attack harnesses and evaluation frameworks) to generate attack variations, run large evaluations, and continuously test AI systems as they change.  Partner with product, engineering, Responsible AI, and detection teams to translate findings into mitigations: better system prompts, input and output filters, grounding controls, agent guardrails, and detections.  Evaluate the effectiveness of AI defenses (detections, safety classifiers, filters, and monitoring) and provide strategic recommendations to close gaps.  Conduct deep research into emerging AI attacker techniques and map them to frameworks such as MITRE ATLAS and the OWASP Top 10 for LLM Applications.  Deliver executive-level briefings, technical reports, and prioritized, actionable