Researcher, Safety - Frontier Assurance

OpenAI•San Francisco, CA

4d•$380,000 - $500,000

About The Position

The Frontier Assurance team strengthens OpenAI’s safety decisions through independent expertise, rigorous assessment, and public transparency. We coordinate external evaluations of frontier models and safeguards, develop approaches such as embedded auditing, and manage the misalignment disclosure process and safeguard testing. Our work helps independent experts scrutinize OpenAI’s safety approach and ensures their findings inform consequential decisions about advanced AI systems. As a researcher on Frontier Assurance, you will build the technical foundations for rigorous assessments of frontier AI risks. You will help design technical assessments with external evaluators and internal teams: identify consequential safety claims to test, define the evidence and methods needed to assess them, and address gaps in coverage. You will work closely with external evaluators and embedded auditors, helping them understand the systems they are assessing and access the models, data, and technical information their work requires. Where direct access is not possible, you will conduct experiments and analyses on their behalf, with methods and results they can meaningfully scrutinize. This role combines research engineering, experimental design, and close technical collaboration. You will move between investigating specific safety questions, navigating complex internal systems, and building reusable tools that make assessments more thorough and efficient. Your work will help turn independent scrutiny into concrete evidence about model capabilities, safeguards, and remaining risks.

Requirements

Have strong software engineering skills and experience building research tools, evaluation infrastructure, or systems for analyzing complex data.
Have practical experience working with modern machine learning systems, including designing experiments, running evaluations, and interpreting results.
Can navigate unfamiliar codebases and infrastructure, debug across system boundaries, and turn ambiguous research requests into reliable technical implementations.
Bring sound judgment to experimental design and measurement, including identifying confounders, limitations, and gaps in evidence.
Enjoy working closely with researchers, explaining technical systems, and helping others investigate difficult questions.
Value independent scrutiny and can support rigorous assessments that challenge internal assumptions.
Have an understanding of, or strong interest in, frontier AI risks, alignment, and the evaluation of safeguards.

Nice To Haves

Experience with AI safety evaluations, adversarial testing, security research, or technical auditing would be especially relevant.

Responsibilities

Help design assessments with external evaluators and internal teams: identify consequential safety claims to test, define the evidence and methods needed to assess them, and address gaps in coverage.
Support the technical execution of third-party assessments and embedded audits focused on frontier risks, working with evaluators to translate research questions into experiments, analyses, and evidence requirements.
Pair closely with external researchers to provide the data, technical documentation, and context needed to assess models and safeguards. Help them understand relevant system behavior, configurations, and limitations.
Design and implement secure access to assessment environments, working with research, engineering, and security teams to make relevant models, tools, and information available.
Conduct experiments and run analyses on evaluators’ behalf when direct access is constrained. Develop methods collaboratively, document implementation choices, and make results reproducible and open to challenge.
Build tooling that automates auditing, assessment workflows, and internal assurance work, including evidence collection, evaluation execution, and tracking changes across model or safeguard versions.
Reproduce external findings, investigate discrepancies with internal results, and work with relevant teams to test proposed mitigations and identify unresolved questions.
Document methods, findings, and limitations clearly so that evaluators and internal decision-makers can understand what the evidence establishes and where uncertainty remains.