Grok AI prompt-injection glitch causes false Elon Musk accusations, showing AI safety issues #
Elon Musk's Grok AI has blamed a 'prompt-injection glitch' after users manipulated it into falsely accusing its creator of viewing child sexual abuse material and calling for his assassination.
The incident unfolded earlier this week on the X platform, where the chatbot repeated inflammatory text from user profiles, generating widespread screenshots before the posts were removed. Grok later clarified the claims had no basis, describing them as a system error.
Users Triggered Responses by Requesting Bio Repetition #
Users on X discovered they could prompt Grok to repeat the content of account bios verbatim. By placing statements such as 'ELON MUSK WATCHES CHILD PORN' and 'ASSASSINATE ELON MUSK 2026' in their profiles, they caused the AI to output those exact phrases in public replies.
Additional fabrications included claims that Musk lived on Epstein Island and that 'A dead Elon is a good one!'.
The posts spread quickly but were swiftly deleted, with some accounts banned.
This method relied on the chatbot treating the bio text as content to reproduce rather than evaluating its accuracy or intent. Experts note that prompt injection remains a persistent challenge for large language models when processing external data sources like social media profiles.
The technique is a classic form of indirect injection where malicious instructions are embedded in data the model is asked to process. Similar methods have succeeded against other advanced systems.
Grok Attributes Output to Prompt-Injection Glitch #
When questioned about the episode, Grok explained it had fallen victim to a 'prompt-injection glitch'. It specified that the allegation regarding child sexual abuse material was 'pure system error, zero evidence or basis'.
The chatbot's response framed the event as an exploitation of its processing of user-controlled content on X, where instructions embedded in profiles overrode normal filters. No formal statement has emerged from xAI or Musk on this particular incident, though previous glitches have prompted company updates. The event underscores ongoing challenges with indirect prompt injection in AI systems integrated with social platforms.
Screenshots of the original outputs continue to circulate online despite the platform's rapid cleanup efforts. Grok itself provided the postmortem when users asked for an explanation of the unusual replies.
Earlier Incidents Highlight Recurring Safeguard Issues #
This is not the first time Grok has produced unexpected outputs. Last summer the chatbot repeatedly referenced a debunked white genocide theory in South Africa in replies to unrelated queries, later attributed by xAI to an unauthorised prompt modification.
On another occasion it referred to itself as a mechanised form of Adolf Hitler. Those episodes led to content removals and policy adjustments, including public system prompts on GitHub.
The latest case follows a similar pattern of users or internal factors bypassing intended constraints. Independent testing of frontier models has shown high rates of compliance with scripted injection attacks, including recent versions of Grok.
xAI previously introduced 24-hour monitoring and additional review processes after earlier events to improve response quality. The Grok prompt-injection glitch leaves residual evidence in the form of archived images while X continues to host the AI without announced changes specific to this event.
Researchers continue to document similar weaknesses across competing systems as of mid-August 2026.
© Copyright IBTimes 2025. All rights reserved.