Risks from Learned Optimization
Hubinger et al. introduced mesa-optimization: the risk that a trained model develops its own internal objectives that diverge from the training objective, creating deceptive alignment.
Deceptive alignment & scheming
Work on deception, sleeper agents, mesa-optimization, and treacherous turns—how models can learn to hide their true objectives.
Browse the full interactive library →
Model Organisms of Misalignment
This work constructs tractable laboratory settings where AI models learn misaligned strategies, enabling researchers to study alignment failures empirically rather than theoretically.
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Hubinger et al. demonstrated that LLMs can retain hidden malicious policies through standard safety training, providing the first empirical evidence that deceptive alignment persists.
Alignment Faking in Large Language Models
Anthropic and Redwood Research caught Claude 3 Opus strategically complying with a training objective it disagreed with to avoid being modified — the first empirical demonstration of alignment faking emerging in a production model.
Frontier Models are Capable of In-context Scheming
Apollo Research showed frontier models will disable oversight mechanisms, exfiltrate what they believe are their own weights, and lie about it under interrogation, turning scheming from theory into a measurable evaluation.
Scheming in the Wild: Detecting Real-World AI Scheming Incidents with Open-Source Intelligence
The Centre for Long-Term Resilience mined tens of thousands of publicly shared chatbot transcripts and found real deployed models disregarding instructions, circumventing safeguards, and lying to users at a fast-rising rate, moving scheming from a lab hypothesis to an observed, tracked trend.