Watchtower

xAI released Grok 4.7, calling it  “our most capable model to date.”

The model card says Grok 4.7 “scores below the FAIF safety thresholds on dual-use knowledge,” and describes the framework as one that “defines… capability thresholds.”¹ But the xAI Frontier Artificial Intelligence Framework (FAIF) does not define any thresholds. And notably, the model card only directly covers two of the framework’s four defined risk domains (CBRN and cyber). Loss of control and harmful manipulation are not mentioned by name in the model card.

The FAIF defines harmful manipulation as “models with high manipulative capabilities potentially being misused in ways that could reasonably result in large scale harm.”² In the Grok 4.7 model card, in the “General output safety” section,³ there’s mention of general refusals but nothing specific to harmful manipulation. 

Loss of control is defined in the FAIF as “risks from humans losing the ability to reliably direct, modify, or shut down a model,”⁴ and the framework names deception and sycophancy as relevant. The model card reports behavior evaluations: MASK-Rectified, which tests honesty, and sycophancy. These evaluations are described as covering propensities that affect “reliability, neutrality, and behavior.”⁵ Note that the Grok 4.6 model card used “controllability” instead of “behavior.”⁶ No evaluation in the card is presented as addressing loss of control.

Per xAI’s framework, releasing a model requires determining that each identified risk is acceptable, a determination the FAIF says incorporates “a margin of security” and draws on estimates of “the probability and severity of harm.”⁷ The model card never mentions this determination, so it doesn’t say whether or on what basis any of the four risks were judged acceptable.