Sean Peters

I asked Sean to put up this proposal and have fully funded his request.

The case for this grant is extremely straightforward: my sense (from ~1.5 years at METR) is that of all the independent external people building directly on top of METR work, Sean is the person whose work my colleagues have been most excited about. Sean appears to be smart, agentic, and focused. My sense is that he's working ~1.5 days per week on these projects basically due to funding constraints, which seems crazy; I'm very excited to support him increasing this to 4.

I could say some more specific things (excited for Sean's connections to pen testers, excited for him to contribute to sabotage monitoring projects, excited for his past work measuring time horizons in offensive cybersecurity , etc.) but mainly my decision comes down to my colleagues' rare degree of excitement for whatever Sean has in store next. I'm thrilled to bet on that.

---

Evaluating Model Attack Selection and Offensive Cyber Horizons, Dec 2025

Evaluation science uses deep statistics. For example, in the attack selection methodology I'm working with hierarchical Bayesian modeling, adaptive sampling, MCMC, etc.

Noting that a lot of stats we use at METR is intentionally pretty simple. E.g. I think we've never done Bayes, despite having many members of staff who would feel intuitively on board with that. Partly this is for communication clarity, partly because it gives less surface area to mess up.

---

Evaluating Model Attack Selection and Offensive Cyber Horizons, Dec 2025

Donating on identical thesis to previous grant -- I wish I could have gone in for more size then!

I haven't thought about expenses for RE; from my perspective, I'm also thinking of this funding as a bet on Sean.