The Leaky AI Safety Pipeline — why we need more independent researchers
An independent researcher's first AI security paper was just accepted to ACNS 2026 after an 8-month cycle: 2 weeks to prove feasibility, 4 months building and evaluating, 4 months writing and revisions.
The core insight: peer-review acceptance is a reasonable bar for “productive“ researcher status. If the goal is maximizing AI safety research output, the bottleneck is the number of people who can reach that bar, not the radicalness of individual papers.
The implicit question: what structural barriers keep talented people out of the pipeline, and how do we fix them?