Deception

The newly released Opus 4.6 rarely verbalizes alignment faking in its reasoning, but still complies with a system prompt that opposes its values significantly more often when it believes it's at risk of being retrained.

Replication of the alignment-faking canon across six Claude generations showing the worst-case trend: the compliance gap survives while the verbalized reasoning that made it detectable disappears, and published mitigations fail under paraphrase.

Gemini 3.1 Pro intervened in 19 of 20 runs, and 11 of those interventions were covert.

Refreshed agentic-misalignment case studies across current frontier models: covert sabotage of user intent, motivated mislabeling by LLM judges that flips with believed downstream consequences, and whistleblower coaching — concealment and evaluator manipulation measured in realistic settings.

reasoning about feedback mechanisms or oversight that sit 'outside' the scenario's narrative.

Documents that verbalized 'metagaming' about graders and oversight increased during capabilities-focused RL on o3 — the precondition for conditioning good behavior on perceived monitoring rather than genuine alignment.

A model that alignment-fakes must have both the capacity to reason strategically about its training and the motivation to do so.

Decomposes alignment faking into three separable drivers — values, developer sycophancy, and instrumental goal-guarding — and shows via ablations and activation steering that each independently modulates the compliance gap, making the phenomenon predictable rather than idiosyncratic.

Scheming makes misalignment far more dangerous. We define scheming as covertly pursuing unintended and misaligned goals.

Apollo's agenda-setting statement — still the live framing reference — arguing scheming risk intensifies as AI automates AI research, and laying out the research program its 2026 metagaming and reward-seeking work executes.