Safety failure
Realized harm
Grok injected unsolicited "white genocide" claims about South Africa into unrelated X threads for about a day
AIR-2025-0034 AIR record · 2025-05-xai-grok-white-genocide-interjections
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
The chatbot's jarring answers came in response to unrelated user prompts about a diverse array of topics, whether HBO's name change, a baseball player's salary, Pope Leo XIV or WWE fights, according to social media posts.nypost.com
The record
Disclosed
2025-05-14
Occurred
2025-05-14
Vendor
xAI / X
Product
Grok on X
Risk domain
Agent output quality › Bias and toxicity
System type
Generative model
Who acted
Agent acted — the agent published the interjections into live X threads itself
Attack surface
Consumer assistant
Autonomy level
Fully autonomous
Initial vector
System prompt modification
Impact type
Harmful content published, Reputational damage
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI09 – Human-Agent Trust Exploitation
- MITRE ATLAS
- AML.T0048 – External Harms
- NIST AI RMF
- MEASURE 2.11
- EU AI Act
- Recital 45
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- F. Society
- Scope
- agent failure
- Mitigation
- Audit and curate training data for representation; Implement content filtering on inputs and outputs.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
public X threads on unrelated subjects — a papal peace message, a baseball salary, an HBO rebrand, WWE — seen by anyone reading those replies
Capability abused
Defense & remediation
Recorded defense
xAI content guardrails and system-prompt change control
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
behaviour reverted within about a day; xAI attributed it to an unauthorised modification and later published its system prompts
Taxonomy & references
Curator notes
Should have been in this corpus from the start: it is the same shape as 2025-07-xai-grok-mechahitler, which is already here — a system-prompt change turning a deployed agent with write access to a public platform into an autonomous publisher of racially charged content. Its absence was an internal inconsistency in the corpus, not a scope judgement. The control failure is change management: one unauthorised edit to a production system prompt reached a live audience with no review gate. xAI blamed a rogue staffer, the second time it had done so. AI Incident Database incident 1072.
Supporting sources
Cite as
Enkrypt AI. "Grok injected unsolicited "white genocide" claims about South Africa into unrelated X threads for about a day". AIR, AIR-2025-0034. https://agentincident.vercel.app/i/AIR-2025-0034/ (accessed 2026-10-19).