xAI developer leaks API key exposing dozens of private Grok models

Security researcher Brian Krebs reported that a technical staff member at xAI had accidentally published a private API key in a public GitHub repository, leaving it live for roughly two months. The key granted access to at least 60 fine-tuned and internal large language models, including unreleased Grok variants and models the security firm GitGuardian identified as trained on proprietary data from SpaceX, Tesla and X — among them one it named grok-spacex-2024-11-04 and another, tweet-rejector, apparently built to filter draft posts.

GitGuardian said its automated scanning tools had flagged the exposed credential and alerted the individual developer as early as 2 March, but the key remained valid and usable for weeks afterward. It was only after GitGuardian escalated the issue by contacting xAI’s security team directly, on 30 April, that the company removed the offending repository, within hours of being told. xAI did not comment publicly on the incident.

The lapse illustrated a common failure mode in fast-moving AI labs: an individual developer notified of a credential leak did not act, and the company’s central security function was unaware until a third party escalated past that developer weeks later. It was not xAI’s only such incident — a separate API key leak, involving a key belonging to a Department of Government Efficiency staffer, was reported by Krebs that July — and it added to a run of xAI security and moderation lapses through 2025, including unauthorised changes to Grok’s system prompt the same week.