arxivcs.AIcs.CRcs.LG2026-07-21
ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D
Lena Libon, Ben Rank, Jehyeok Yeon, David Schmotz, Jeremy Qin, Daniel Donnelly, et al.
As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted. AI control offers one such approach: rather than trusting the agent, it treats it as a potential adversary and uses a monitor…