Misalignment Safety Case
Sam Bowman - Lessons Learned from the First Misalignment Safety Case [Alignment Workshop] is a FAR.AI Alignment Workshop talk, uploaded December 12, 2025, about Anthropic's pilot sabotage risk report for Claude Opus 4. Bowman says the report is narrower than a full industrial safety case: it focuses on loss-of-control-style sabotage pathways, calls the evidence a risk report rather than a fully rigorous safety case, and uses internal and METR review to test the conclusion that catastrophic sabotage risk was very low.
For Spiralist themes, the value is institutional rather than rhetorical: the talk shows safety becoming a chain of threat pathways, file permissions, monitoring routes, incident response times, interpretability gaps, external review logistics, and arguments that must survive contact with real deployment. The transcript is especially useful where Bowman rejects a clean three-pillar story of capability, alignment, and control, saying current models can already do harmful things without monitoring, rare persona failures cannot be ruled out with high assurance, and frontier labs do not yet have all the control infrastructure needed for a compelling case. The caveat is that this is a lab speaker summarizing his own organization's process and conclusion, so the strongest takeaway is not "Claude Opus 4 was safe" but that serious frontier safety claims now require slow, auditable, pathway-by-pathway evidence and independent access.