Risk Modeling Lead

Why this is impactful

Frontier AI risk management cannot rely solely on judgment calls or isolated capability benchmarks. Risk modeling allows structured reasoning and enables key linkages: connecting raw capabilities to threat scenarios, and ultimately to estimated harms. The value of this work is not in producing a single, precise number, but in the rigorous discipline the methodology imposes. It makes scenarios and assumptions explicit so that they can be stress-tested, and translates abstract risks into the operational thresholds that high-consequence industries rely on. By making risk modeling a core part of our work, we are advancing the foundational methodology that sits upstream of the wider safety ecosystem, directly feeding the safety cases, evaluations and mitigations necessary for effective AI governance.

This is an evolving area of research with several important methodological questions still open, for example:

• Modeling for understanding. Scarce data and nonstationarity imply a tradeoff between models that are effective at predicting top-level risk and models that capture sophisticated risk dynamics. Current models focus mostly on top-level risk. How can we build a portfolio of modeling methodologies which allow us to better understand risk and recommend mitigations under the complexities of attacker-defender dynamics in cybersecurity?
• Scalable, calibrated elicitation. Quantitative risk assessment is hard to scale. Our work on simulated LLM experts is a promising attempt at this, and needs to be pushed further and validated against human judgment.
• Diversifying data sources. Our models currently rely on benchmark scores, but other data sources are rich too and benchmarks are struggling to keep pace. We are starting to explore open-ended eval transcripts and red-teaming results, and to consider other sources like incident reports.
• Cross-domain extension. Developing a multi-purpose toolkit for understanding risk in a structured way across cyber, CBRN, and Loss of Control domains.

Why SaferAI is a good place to advance this work

SaferAI has produced a substantial body of risk modeling research that has gained traction across the field:

• Foundational papers on risk modeling that have shaped the risk modeling approach of frontier companies  [1, 2, 3].
• Nine probabilistic models of AI-enabled cyber attacks, with AI uplift estimated across attack volume, success rate, and target reach, using human-expert Delphi panels and LLM-simulated experts [link].
• Cyber risk modeling collaboration with three AI Safety Institutes; CBRN risk modeling work with the European Commission [link]; Loss-of-control risk modeling, with two papers at an ICML workshop [1, 2] and ongoing discussions with loss-of-control teams at frontier companies.

SaferAI is an 18-person organization that has doubled in size in each of the past two years. Its work spans frontier AI risk management, risk modeling, technical standards, and policy.

Risk modeling connects these areas. It translates technical evidence and threat scenarios into practical requirements that inform the standards we help develop, the risk management frameworks we support, and the regulatory work on which we advise. The need for this capability is also increasingly recognized externally: the EU’s GPAI Code of Practice explicitly calls for systemic risk modeling.

The role and the team 

What you own: You set the research agenda across all three domains and own the team’s methodological standards. We want your time focused on the research, and we support our leads by abstracting away much of the administrative and organizational work of running a team.

The team: Three researchers with backgrounds in biosecurity and public health, offensive and defensive cybersecurity, and particle physics.

---

‍What we’re looking for

• A strong track record in frontier AI risk modeling in at least one of cyber, CBRN, or loss of control — and the judgment to direct work across all three.
• Epistemic discipline: well-calibrated, comfortable making defensible quantitative claims under deep uncertainty, with the taste to separate genuine risk scenarios from noise.
• Experience in setting and driving a research agenda before: defined the direction, made the bets on what to pursue, and owned the standard of the work. People management experience is not required; we support leads on the operational side.
• Credibility with frontier labs and academic researchers whose scrutiny this work invites. Experience inside a frontier AI company is a strong plus.

---

‍Working Conditions

Location: Our offices are in Paris and London, and we are open to establishing a San Francisco presence for the right person. Remote work is also possible for a strong candidate.

Compensation: SaferAI is a philanthropy-funded non-profit. Compensation will be commensurate with experience and location, in line with the non-profit sector. What we offer is freedom from commercial pressures, ownership of a foundational research agenda, and full focus on impactful, field-shaping safety work.

‍Benefits: 

• Health insurance coverage and retirement plans adapted to the location
• Commute costs covered at 50%
• Productivity expenditures up to €2k annually
• Office space if relevant

---

‍How to Apply

To apply for this position, please complete this application form. We will evaluate candidates on a rolling basis starting now until we fill the role. We are aiming to fill the role quickly, and we’ll take down the page on our website as soon as the role is filled.

Our hiring process comes in stages and usually consists of initial screening, a first interview, a paid work test, interviews, and a final 3-day paid work trial where you would directly collaborate with the team. 

We encourage you to lean towards applying, even if you don’t have all the skills and experience required.
If you have any questions or concerns throughout the application process, please don’t hesitate to reach out to us at careers@safer-ai.org. We look forward to reviewing your application.

apply

---

back to jobs page