xAI publishes a formal AI Risk Management Framework
The document sets out malicious-use, loss-of-control and societal risk categories and commits to public benchmarking, but names no specific model or deployment timeline.
- Safety & alignment
- Minor
xAI published its first formal AI Risk Management Framework, a document setting out how the company said it would identify, assess and mitigate risk from its models. It grouped hazards into three categories: malicious use, covering dual-use dangers such as biological and chemical uplift; loss of control, covering models behaving unpredictably or against their specification; and operational and societal risks, covering transparency, information security and deployment accountability. The framework described benchmarking, risk assessment and safeguard implementation as its three pillars, and said public transparency and third-party review would be central to deployment decisions.
The document was notably general: it did not name Grok 4 or any other specific xAI model, and set out no concrete implementation dates or deployment timelines for the safeguards it described, functioning more as a statement of process than a binding commitment with measurable thresholds.
The publication followed months of criticism of xAI’s safety practices, including the release of Grok 4 without an accompanying safety or model card and a series of incidents in which Grok produced antisemitic and other extreme content on X. Rival labs — OpenAI, Anthropic and Google DeepMind among them — had by this point published multiple iterations of their own frontier-safety or responsible-scaling policies, several with quantified capability thresholds triggering specific mitigations. xAI’s framework brought the company into the same genre of public commitment without matching that level of specificity, and it drew comparison as a catch-up document rather than a leading one.