About Dune Science

Mission

Frontier AI developers increasingly rely on safety cases to justify training and deploying their most capable systems. A safety case is a structured argument, backed by evidence, that a system is safe enough to use in a given setting [1, 2], and developers have begun to publish them [3, 4, 5]. But a safety case is only as strong as the evidence beneath it, and today that evidence is produced and judged almost entirely by the developer making the case.

In nuclear power and aviation, operators do not certify their own safety: independent assessors test the case before a system is allowed to run. The safety case literature has called for the same in AI [1, 2, 6]. Dune Science exists to provide that check, by establishing through reproducible experiments which claims in AI safety cases actually hold.

What we do

We test published safety cases empirically. Most proposals, whether based on AI control, chain-of-thought monitoring, scheming evaluations or debate [7, 8, 9, 10], are still sketches: careful arguments resting on assumptions that have never been tested in a realistic setting.

For each safety case, we build an open-source, end-to-end implementation of the system it describes, break the argument down into the individual claims it depends on, and design experiments that try to falsify each one. We then publish a claim-by-claim report of what held, what failed and under which assumptions, with the code, logs and data needed to reproduce every result.

This work gives decision-makers evidence on frontier AI safety from an organisation with no stake in the verdict. It shows researchers and funders which parts of a safety case are weakest, and therefore where research is most needed. And it leaves the field with open implementations that developers, AI Security Institutes and academics can rerun and extend. We publish negative results as readily as positive ones, and we do not accept funding from frontier AI developers.

Team

• AK Aton Kamanda Research Engineer, LawZero
• LM Lucas Maes PhD student, Mila & Université de Montréal

Read our full approach →