Researcher, Recursive Self-Improvement Safety

Written by the T&S Network from OpenAI's posting.

Preparedness team work on frontier AI risk, focused on loss of control and recursive self-improvement. The role builds mitigations ahead of accelerated AI development, covering pre-deployment risk assessment, control measures, RSI-relevant training interventions, and the institutional practices and external communications that come from that technical work. The team measures and predicts frontier capabilities, maintains mitigation targets under OpenAI's preparedness framework, and partners with other staff to meet them.

• AI misuse and jailbreaks

What you'd do

• Anticipate future misalignment threats and decide which problems to prepare for first.
• Turn broad risk goals into concrete research directions, such as stress-testing monitors for scheming.
• Build scrappy prototypes and iterate them into established components of safety pipelines.
• Design and run experiments on reward hacking, sandbagging, and scheming, including model organisms of misbehavior.
• Build automated auditing that sifts production traffic for severe misalignment and elicits tail risks before deployment.
• Secure buy-in from other OpenAI staff and communicate work clearly.
• Collaborate with or manage other staff as the team scales to meet these problems.

What it asks for

• Strong technical execution ability.
• Strong strategic and research judgment, including prioritizing in domains with weak feedback loops.
• Prior work in ML research, AI alignment, or AI verification (preferred).

Terms in this role

Red teaming.

Testing a system, often an AI model, by deliberately trying to make it fail or misbehave, so problems are found and fixed first.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on OpenAInorth_east before you apply. If we got something wrong, report this listing.