Artificial Intelligence Researcher

Space Executive; United States; On-site; USD 180,000 - 250,000; Full time

Job summary

Space Executive is seeking a leader to define how the industry measures AI risk and to own the taxonomy, evaluation harness, and final releases of each benchmark. You will work with internal specialists and external experts, ensuring rigorous standards while deciding what is published openly versus shared privately with labs and developers.

The role reports to the CTO and collaborates with the research lead on public research direction.

Qualifications

• Masters or PhD in CS/ML or equivalent research depth.
• At least three years designing safety evaluations for LLMs in production.
• Proven publication record in AI safety and security (5+ papers; 2 first authors).
• Strong engineering: build eval harnesses, distributed inference, and familiarity with codebases.
• Ability to design a taxonomy from scratch and lead researchers/freelancers.
• Excellent written and verbal English; ability to coordinate across time zones.

Responsibilities

• Deliver benchmarks with a fast release rhythm, roughly every 2–3 weeks.
• Own taxonomy, evaluation harness, quality standards and final releases.
• Coordinate with internal specialists and external subject-matter experts as needed.
• Manage plan, schedule, and freelance contributors to keep projects on track.
• Shape the roadmap with CTO and leads; review research with customers and the broader landscape.

Skills

English proficiency

Research leadership

AI risk evaluation

Harm domain knowledge

Education

Masters/PhD in CS/ML

Tools

vLLM

Python

Job description

This is a role for someone who wants to define how the industry measures AI risk. Every few weeks, you will release a new benchmark targeting a frontier risk area that hasn't yet been properly quantified. Depending on sensitivity, some of this work is published openly, while some is shared privately with model developers. A number of projects are run jointly with major AI labs and academic partners.

You won't be building every eval on your own. For each benchmark, you'll be paired with an internal specialist who owns the relevant harm domain, and you'll have budget to bring in and direct external subject-matter experts. Your ownership covers the taxonomy, the evaluation harness, the quality standard and the final release.

The position reports into the office of the CTO and works closely with the research lead responsible for the company's public research direction. You'll also be able to draw on a large in-house bench of harm-focused researchers whenever a topic calls for extra depth.

What You'll Be Doing

Delivering on a fast release rhythm. Expect to put out a new benchmark roughly every fortnight to three weeks. Scope flexes with the topic: a conversational taxonomy might hold around a hundred evals, whereas an agentic or RL-based benchmark may be closer to twenty, given how costly each item is to review. How sensitive the material is will determine whether it goes public or stays with the labs.

Holding the standard. Success looks like this: a frontier lab re-runs your benchmark and reproduces your results. Verifiers are robust, rubrics are unambiguous, the data distribution makes sense, and the lab's own domain experts see the taxonomy as genuinely new. Model-assisted screening is fine, but you'll still go through the evals personally, catch anything that drifts from the taxonomy, and send it back when it isn't right.

Keeping delivery on track. You own the plan and the schedule, and you're the one making sure contributing researchers hit their dates. When needed, you'll bring in and steer a small number of freelance experts directly.

Shaping the roadmap. Around once a month, you'll join the CTO alongside pod and research leads to review what the research teams are seeing, what customers want, and what's happening across the wider landscape. The outcome is an updated quarterly release plan, aligned with the commercial relationships the business is looking to build.

Staying plugged into the field. Roughly a fifth of your time is set aside for the wider ecosystem: keeping up with new research, maintaining relationships with people inside the labs, understanding what's worrying them, and attending a handful of conferences each year. Regular, ideally weekly, conversations with lab contacts are part of the job.

What We're Looking For

• A Masters or PhD in computer science, machine learning or a similar discipline, or comparable depth gained through industry research.
• At least three years designing and operating safety or security evaluations for LLMs in production, whether at a frontier lab, a model provider or a specialist safety/security research group.
• A track record of publishing in AI safety and security, with five or more relevant papers and first authorship on at least two.
• Solid engineering skills: building eval harnesses, distributed inference, vLLM, and the ability to dig into an unfamiliar codebase and fix it.
• The ability to design a taxonomy from scratch, not just evaluate against an existing one.
• Comfort leading a researcher and a couple of freelancers through influence rather than formal line management.
• Excellent written and verbal English; much of the work depends on clear communication across time zones.
• Genuine curiosity about the harms being studied, and an appetite for getting up to speed on a new domain every few weeks.

Bonus Points

• Post-training experience across SFT, DPO or GRPO, particularly reward design for subjective or safety-critical objectives.
• Experience evaluating agentic systems, including tool use, orchestration, permissioning and prompt injection.
• Papers accepted at leading ML or security venues.
• Confidence presenting your research to customers and to large or senior audiences.
• Happy to travel to three or more conferences a year.

About the Company

Our client is an established AI safety and security business working with many of the world's leading frontier labs. Its platform protects the technologies people rely on to communicate and create, whether they're interacting with one another or with AI systems.

The company covers the full model lifecycle, from pre-release hardening and red-teaming through to live guardrails and ongoing monitoring, serving frontier labs, enterprises and major consumer platforms. Its team is recognised as one of the strongest in the space, and its work protects users at very large scale.

If you're driven by the challenge of making AI safer and want your research to directly shape how frontier models are tested, we'd love to hear from you.