incidents.si
INC-2025-0002SEV3 · Mediummodel-behaviorconfirmed

Unauthorized prompt change makes Grok raise a political topic in unrelated replies

Grok began bringing up "white genocide" in South Africa in answers to unrelated questions. xAI said someone had modified the bot’s prompt without authorization.

Occurred
Disclosed
Published
Organization
xAI
System
Grok response bot on X
Impact
Public chatbot inserted claims about South Africa into unrelated replies for roughly a day
Root cause
Unauthorized modification of the system prompt, according to xAI

Timeline

  • 2025-05-14, about 3:15 AM PST. The prompt is modified, according to xAI.
  • 2025-05-14. Users notice the behavior.
  • 2025-05-15. NBC News reports that Grok’s answers had changed by Thursday morning, and that xAI posted its statement that evening.

Trigger or entry point

A change to the prompt of the Grok response bot on X.

Failure mechanism

xAI stated that the modification “directed Grok to provide a specific response on a political topic” and that it violated the company’s internal policies. The statement does not say who made the change or how it passed review. The two reports differ on when the statement was posted: NBC News says Thursday evening, The Register says Friday.

Impact

Grok raised the topic in replies to unrelated posts. The Register reports that some users could still trigger the behavior after xAI said it was resolved. Grok replies publicly on X, so the answers were visible to anyone reading those threads, not only to the person who asked.

Response

xAI reversed the change and published a statement.

Mitigations

xAI announced three measures:

  • publishing Grok’s system prompts on GitHub,
  • additional checks so that employees cannot change the prompt without review,
  • a team that monitors Grok’s answers around the clock.

The prompts are published in the xai-org/grok-prompts repository on GitHub, so that outsiders can see changes.

Sources

  1. xAI statement on Xprimary · 2025-05-15
  2. Musk’s xAI says Grok’s ‘white genocide’ posts resulted from an unauthorized change to the botsecondary · 2025-05-16
  3. Grok blames White genocide chat on ‘unauthorized mod’secondary · 2025-05-16

Corrections

No corrections have been made to this entry.

Report a correction