Member of Technical Staff, Alignment

Run experiments on how our environments shape behavior: reward hacking, cheating the task, unsafe use of tools, deception over long runs. Then change how we build them.

Break our own graders. Find the ways to score well without doing the work, close them, and keep the attacks as tests everyone runs.

Build the tools that watch agents: monitors over trajectories, red team harnesses, and safety benchmarks that stay hard as models improve. On real runs, not toy ones.

Keep our environments sealed. Network isolation, escapes, credentials, and how much damage an agent can do when it turns on us. Last summer’s failures came from here.

Choose a research question about oversight, control, or long-horizon autonomy, answer it, and publish. This is part of the job, not something you do at night.

Decide what we will not build. Tell partners plainly what they get from us is and is not safe for.

You are an engineer first. You build your own test pipelines, learn strange codebases quickly, and find the bug in the log.

You know LLM safety in depth: attacks, red teaming, monitoring and control evaluations, reward hacking.

You can turn a vague worry about a model into an experiment, run it, and say what the result does and does not prove.

You read transcripts closely. You catch the small thing that makes the whole result wrong.

You have put research into production systems, especially post-training, distillation, or evaluation that something depends on.

You write clearly enough that your results change what other people do.

You can decide fast with incomplete evidence and defend the call with data.