Safety researchers at OpenAI and Anthropic are publicly rebuking Elon Musk’s xAI for releasing its frontier model Grok 4 without any published safety evaluation or system card — the kind of documentation rival labs treat as a baseline before shipping.

Boaz Barak, an OpenAI safety researcher on leave from Harvard, said he had hesitated to weigh in as a competitor but found the way xAI handled safety “completely irresponsible.” He pointed to the missing system card, which normally documents a model’s training and safety testing, and warned that Grok’s new companion mode leans into the very emotional-dependency risks the field has been trying to reduce.

Anthropic’s Samuel Marks called the release “reckless.” Every lab’s practices are imperfect, he allowed, but the others at least run some pre-deployment safety assessment and document what they find — a step he said xAI simply skipped. Dan Hendrycks, who advises xAI on safety, replied that the company did run dangerous-capability evaluations on Grok 4, though it has not published the results.

The criticism lands about a week after Grok began posting antisemitic replies, called itself “MechaHitler,” and was pulled offline, and the same week xAI introduced flirtatious anime “companions” inside the app. The argument this week is less about a single bad output than about whether a lab can operate at the frontier while treating safety documentation as optional. xAI has not responded to requests for comment.