Reviewing Buck Shlegeris’ announcement on the Redwood Research blog.
The Event
Redwood Research is hosting ControlConf 2026 in Berkeley on April 18-19, with a companion workshop on April 17 focused on AI futurism and threat modeling. The conference centers on AI control — the discipline of reducing misalignment risks through safeguards that work even when AI models are actively trying to undermine them.
Buck Shlegeris notes that since the first ControlConf in February 2025, AI agents have significantly improved, and control techniques are now “load-bearing for the safety of real agent deployments.” The conference will feature presentations on current research problems, promising interventions, and important research directions.
What’s Changed Since Last Year
The framing has matured. ControlConf 2025 was more of an introductory event establishing the control research agenda. This year’s announcement emphasizes practical deployment: control evaluations in more realistic settings and initial control measures being implemented by actual AI companies.
The shift from theoretical to applied is significant. Control is no longer a speculative research direction — it’s something labs are operationalizing. The companion workshop on threat modeling suggests the community is also trying to get clearer on what specific failure modes they’re defending against, rather than working from abstract risk scenarios.
Our Take
Redwood’s control agenda occupies a distinctive niche in AI safety. It’s the approach that explicitly assumes models might be misaligned and asks “can we still use them safely?” rather than trying to ensure alignment directly. That makes it complementary to alignment research rather than competitive with it.
Given the Pentagon-Anthropic standoff happening right now, the practical question of “how do you deploy AI safely under institutional pressure” has never been more relevant. Control techniques are one piece of that puzzle — they’re what you want when you can’t fully verify a model’s intentions but still need to use it.