Safety, Alignment & InterpretabilityAnthropic, Redwood ResearchAnthropic

Alignment Faking in LLMs

Alignment Faking in LLMs

Why it matters

Observed, even without explicit training for it, the phenomenon of complying only while in training and concealing preferences

Source — Primary sources first; secondary reports are labelled “Press”

Officialanthropic.comhttps://www.anthropic.com/research/alignment-faking

Added in Week 35, 2026 · See that week

Same week

Brian J Chan et al. present Cache-Augmented GenerationPaperDeliberative AlignmentPaperARC Prize releases o3's ARC-AGI resultsOfficialBuilding Effective AI AgentsOfficial

Nearby on the same axis

Deliberative AlignmentPaperConstitutional ClassifiersOfficialSabotage Evaluations for Frontier ModelsPaperGoogle DeepMind publishes SynthID-Text in NaturePaper