Reinforcement Learning Engineer

Role overview

Build reinforcement learning environments that let AI agents practice cybersecurity tasks and receive rewards based on verifiable execution outcomes. The role connects vulnerability research and low-level security analysis with language-model reasoning, while developing reliable infrastructure for large-scale, repeatable evaluation and training.

Responsibilities - Design and deploy high-throughput environments that assess agent actions against deterministic results. - Create test harnesses, sandboxes, and verification systems that turn vulnerability challenges into clear pass-or-fail rewards. - Connect automated security analysis and vulnerability data to pipelines that generate training environments. - Build isolated, reproducible execution setups for parallel exploitation and remediation runs. - Develop telemetry, ground-truth checks, and trajectory logs to study agent behavior and detect reward gaming or false positives. - Instrument and debug memory, process, and network behavior; improve reset speed and infrastructure capacity for distributed training. - Benchmark AI models on tasks such as fuzzing, exploit generation, and patch validation, and collaborate with external research teams on evaluation formats.

Requirements - Understanding of reinforcement learning workflows for language models, especially execution-based feedback or verifiable rewards. - Strong Python development skills and systems programming experience in C. - Solid knowledge of software vulnerabilities, binary exploitation, fuzzing, or program analysis. - Experience with Linux, low-level debugging, DevOps pipelines, and reproducible build or container tools such as Docker, BuildKit, or Nix. - Experience building or using benchmark environments, security challenges, or execution sandboxes.

Nice to have

Rust experience and experience designing custom reward functions or security challenge environments are useful. The work spans offensive and defensive security evaluation, with emphasis on deterministic verification, safe isolation, and scalable execution.