The role
- COMP
- $250K - $450K
- EQUITY
- 0.5 - 1.5%
- LOCATION
- San Francisco
- WORKPLACE
- On-site
- EXPERIENCE
- 2 - 12 years
- VISA
- None, Visa transfers
- STACK
- Python, PyTorch
- INDUSTRY
- AI, Cybersecurity, Security, Software Development
/roles — ROLE_328
Member of Technical Staff - Research (Cybersecurity/Benchmarking)
Small AI security lab training models for adversarial testing, red teaming, and runtime defense
The company
AI security research lab building red-teaming, adversarial simulation, evaluation, and runtime protection tools that find how models fail and make them more robust.
- STAGE
- Series A-stage
- FUNDING
- $10M+ raised
- TEAM
- founding team of <10
- FOUNDED
- 2024
JD — the work
About the role
You would post-train models with reinforcement learning to give them adversarial capabilities, working across the full training pipeline, covering environment design, generating data, and evaluation. The work mixes distributed systems engineering, new methods for settings where data is scarce and signal is noisy, and algorithmic improvements in how trained models are used for adversarial simulation. The lab is a small, in-person San Francisco team spanning product, engineering, and research.
What you'll do
- Post-train models with RL to build adversarial capabilities for red teaming and simulation
- Own the training pipeline from environment design and data generation through evaluation
- Solve the distributed systems challenges that come with large-scale RL training
- Invent methods that work in data-constrained, low signal-to-noise settings
- Make algorithmic gains in how trained models are used for adversarial simulation
- Build benchmarks and evaluations that measure the cybersecurity capabilities of models
What they're looking for
- Academic or industry background in reinforcement learning or ML research, or a strong drive to learn it
- Comfort with PyTorch and RL or post-training infrastructure, such as verl or Tinker
- Security experience, or real excitement to build it on the job
- High agency, strong technical judgment, and clear thinking
- Results focus: able to go deep on exploration while still shipping practical implementations
Nice to have
- Publications at top ML venues (ICML, NeurIPS, ACL, ICLR, CVPR, or similar)
- Excitement about picking up new domains and going deep on AI security
- Willingness to step outside your remit to unblock teammates