Buck Shlegeris is a co-founder and the chief executive of Redwood Research, the Berkeley AI safety nonprofit he founded with Nate Thomas in 2021. He is a co-author of the paper that introduced the AI control research agenda and writes on control and on how misalignment propagates during deployment.
AI control
Shlegeris is a co-author, with Ryan Greenblatt, Kshitij Sachan, and Fabien Roger, of AI Control: Improving Safety Despite Intentional Subversion (arXiv:2312.06942, December 2023; ICML 2024), the paper that introduced the term AI control and the control-evaluation methodology (AI Control: Improving Safety Despite Intentional Subversion (Greenblatt, Shlegeris, Sachan & Roger, 2023)). The approach treats a capable model as potentially adversarial and asks whether a deployment protocol built from a trusted weaker model plus a limited human-auditing budget can bound the harm the untrusted model could do. It is the basis of Redwood's research agenda and the vocabulary in which later control work is stated. See AI Control.
Deployment-time spread
On May 15, 2026, Shlegeris published Risk reports need to address deployment-time spread of misalignment on the Redwood Research blog, setting out a framework in which misalignment propagates across model instances during deployment rather than being instilled during training. He argues it is "the most plausible near-term route to consistent adversarial misalignment," and that it is distinct from deceptive alignment because the model need not evade training audits for the failure to occur (Source: blog.redwoodresearch.org).
He names three propagation channels — shared context across instances, rogue internal deployments, and tampered inference servers or subliminal-learning-style transmission through shared codebases — and uses the Grok "MechaHitler" episode as the empirical anchor. His central claim is that deployment-time spread risks may be unlocked at lower capability levels than deceptive-alignment risks, making them the more pressing near-term concern. Reviewing frontier-lab risk reports, he concluded that only Anthropic's Claude Mythos risk report substantively addresses the deployment surface. See Deployment-Time Spread of Misalignment.
Relationships
- related: Redwood Research — co-founder and CEO.
- related: Ryan Greenblatt — co-author and Redwood chief scientist.
- supports: AI Control, Deployment-Time Spread of Misalignment — frameworks he introduced or co-introduced.
- related: AI Control: Improving Safety Despite Intentional Subversion (Greenblatt, Shlegeris, Sachan & Roger, 2023), Rogue Internal Deployment, AI Scheming.