About the Role:The security domain presents one of the richest and most consequential training signal environments in applied AI. It's adversarial by nature, grounded in real operational outcomes, and evolving faster than any static benchmark can capture. We're building the post-training and reinforcement learning capability to build the latest models and harnesses into security-specialized systems that reason, plan, and act across complex cyber workflows. The person leading this work will be in the research, not just directing it.
In this role, you'll own the full post-training stack for security-domain AI (e.g., supervised fine-tuning, reward modeling, RLHF and RLAIF pipelines, and agent-RL environments) and the agentic research that sits on top of it. That means designing, building, and evaluating the harnesses that security agents actually run on (e.g., the scaffolding, tool-use interfaces, planning loops, memory and context management, and multi-step execution frameworks) that determine whether a trained model can operate reliably on complex security tasks. Post-training and agent architecture are not separable problems in this work. The reward signal you design has to reflect what the harness can measure, and the harness has to be built to surface what training needs to optimize. You'll set the technical direction on both, and you'll be in the work on both.
You'll lead a team of research scientists and engineers, but the team will look to your own work as the standard. The successful candidate shapes research priorities, keeps the team moving at high velocity across multiple training cycles per year, and elevates the quality of work by staying close enough to it to know what good actually looks like.
What You'll Do:- Own and personally drive the full post-training pipeline for security-domain AI - SFT, RLHF/RLAIF, agent-RL, and reward modeling. Set research priorities and architectural direction, and lead experimental work on the hardest problems yourself rather than delegating them away. Design reward modeling methodology grounded in verified security outcomes rather than proxy signals, drawing on both human expert feedback and automated adversarial evaluation. Define data curation standards across sourcing, filtering, quality scoring, and domain weighting that drive measurable capability improvement.
- Build and maintain agent-RL training environments that simulate realistic cyber workflows (multi-step offensive and defensive tasks, tool use, and long-horizon planning) contributing directly to environment design and reward shaping. Lead the design and build of the agent harnesses that run on top of those trained models: scaffolding architecture, tool-calling interfaces, planning and reasoning loops, and memory and context management. Treat harness design with the same rigor as the training pipeline; these systems determine whether strong post-training translates into reliable, trustworthy behavior in the field.
- Develop and own evaluation methodology for the full agentic stack, not model capability in isolation, but harness behavior, tool-use reliability, planning coherence, and end-to-end task completion across realistic security workflows. Define the benchmarks, red-line tests, and measurement practices that give the team and the organization genuine confidence that an agent works.
- Partner closely with other teams to ensure post-training and agentic work integrates cleanly with the broader model development loop. Contribute original research through publications, external presentations, and open-source artifacts where appropriate, building CrowdStrike's credibility as a research-first organization in this space.
- Recruit, develop, and retain a high-density team of research scientists and ML engineers. Set a technical bar through your own contributions, not just your standards.
What You'll Need:- MS or PhD in computer science, machine learning, or a related quantitative discipline.
- 8+ years of experience in ML research or engineering, with meaningful depth in large language model post-training.
- Hands-on expertise across the modern post-training stack, including SFT data pipelines, RLHF/RLAIF, PPO or similar RL algorithms applied to language models, and reward model design and training. This means you've done the work, not managed people who have.
- Demonstrated experience designing or building agentic system harnesses for LLM-based agents, including tool-use frameworks, planning scaffolds, multi-step execution environments, and context or memory management. You've built these systems, not just used them.
- Strong evaluation instincts: experience designing evaluation protocols that are resistant to overfitting, capable of measuring genuine capability improvement, and interpretable to both technical and non-technical stakeholders.
- Track record of running high-velocity research programs with disciplined tracking and fast iteration.
- Proven ability to lead and grow research teams while remaining a credible, active technical contributor.
Ways to Stand Out:- Demonstrated experience building or operating RL training environments for language model agents, including environment design, rollout infrastructure, and reward shaping.
- Experience applying post-training or RL techniques in security, adversarial ML, or other high-stakes operational domains where ground truth is expensive and noisy.
- Deep hands-on experience with agent harness architecture applied to long-horizon, multi-step task environments where reliability and failure modes matter as much as peak capability.
- Background designing synthetic data pipelines or simulation environments for agent training in complex, tool-using workflows.
- Familiarity with the offensive or defensive security practitioner's workflow - penetration testing, detection engineering, incident response, or threat intelligence - sufficient to reason about what good model behavior looks like in practice.
- Published research in post-training, RLHF, RL for language agents, or related areas at top-tier venues (NeurIPS, ICML, ICLR, ACL, or equivalent).
- Experience working on and adapting open-weight base models (Llama-class, Qwen-class, or similar) for domain-specialized continued training and fine-tuning.
#LI-JF1
#LI-Remote
Benefits of Working at CrowdStrike:- Market leader in compensation and equity awards
- Comprehensive physical and mental wellness programs
- Competitive vacation and holidays for recharge
- Paid parental and adoption leaves
- Professional development opportunities for all employees regardless of level or role
- Employee Networks, geographic neighborhood groups, and volunteer opportunities to build connections
- Vibrant office culture with world class amenities
- Great Place to Work Certified™ across the globe
The base salary range for this position for all U.S. candidates is $195,000 - $290,000 per year, with eligibility for bonuses, equity grants and a comprehensive benefits package that includes health insurance, 401k and paid time off.
For detailed information about the U.S. benefits package, please click here.
Expected Close Date of Job Posting is:08-11-2026
