Member of Technical Staff, AI Security Research

About the role

You will research security failures in AI agents, including prompt injection, unauthorized data access, and misuse of tools. The work spans behavior classification, model evaluation, and controls on agent actions and permissions.

You will take a research question through dataset design, experiments, and reproducible results, then work with engineering to test the method in Reagent. Training experiments use synthetic and public reference data, hand-labeled by internal and third-party personnel. Customer data is not used to train shared models.

What you will do

• Build datasets and evaluations for prompt injection, data exfiltration, unauthorized tool use, and other agent security failures
• Fine-tune behavior classifiers and embedding models, and compare them with simpler baselines
• Measure false positives, detection recall, calibration, latency, and inference cost on held-out workflows
• Design and test controls for tool calls, data access, and agent permissions with product engineers
• Document methods, reproduce results, and publish findings with their limitations

What you bring

• Experience training or evaluating machine learning models, with work on language models, embeddings, or agent systems
• Strong Python skills and experience implementing experiments in PyTorch or a comparable framework
• Experience designing evaluations that account for data leakage, label quality, and distribution shift
• Able to read research, implement a method, and explain what the evidence does and does not establish

Additional experience

• Research on prompt injection, adversarial machine learning, or agent security
• Experience with contrastive learning, retrieval, or calibrated classifiers
• Published research, open-source implementations, or contributions to evaluation datasets