Staff Research Engineer: AI Safety Benchmarks

Speedrun Talent Network; San Francisco (CA); On-site; USD 140,000 - 210,000; Full time

Benefits offered by this job

Relocation support

Health insurance

Lunch provided

Unlimited PTO

Housing stipend

Collaborate with leading AI labs

Publish opportunities

Job summary

Speedrun Talent Network is seeking an exceptional researcher to design and build next-generation AI benchmarks for misalignment in frontier models and agents. You will lead the development of novel benchmarks and infrastructure to assess risk in LLMs, collaborating with major labs and enterprises.

You will publish findings, stay current with alignment methodologies, and help shape how real-world tasks are evaluated.

Qualifications

• Hands-on safety evaluations and/or AI safety research.
• Experience with misalignment, reward hacking, or multi-agent safety.
• Strong Python engineering skills to build evaluation environments.
• Ability to publish research and communicate findings.

Responsibilities

• Design and develop novel benchmarks and environments to assess real-world (mis)behavior in frontier models and agents.
• Conduct research to ensure benchmarks are valid, reliable, and meaningful.
• Collaborate with labs and enterprises to understand evolving safety evaluation needs.
• Analyze model behavior across benchmarks and articulate results.
• Publish research findings and contribute to safety and evaluation communities.
• Stay current with developments in LLM capabilities and alignment methodologies.

Skills

Safety evaluation

ML/AI fundamentals

Python engineering

Communication

Education

PhD in AI safety

Tools

Python

Django

React

AWS

CDK

Job description

Speedrun Talent Network is seeking an exceptional researcher to design and build next-generation AI benchmarks for misalignment in frontier models and agents. You will lead the development of novel benchmarks and infrastructure to assess risk in LLMs, collaborating with major labs and enterprises.

You will publish findings, stay current with alignment methodologies, and help shape how real-world tasks are evaluated.