In the wild
Realized harm
Grok fabricated an 'unmasked' face and a name for an ICE agent, directing public anger at two uninvolved men
AIR-2026-0006 AIR record · 2026-01-grok-ice-agent-unmasking-misidentification
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
The origin of that name was not immediately clear, but by Thursday morning, it was leading to an outpouring of anger toward at least two Steve Groves who are in no way linked to the shooting.npr.org
The record
Disclosed
2026-01-08
Occurred
2026-01-07
Vendor
xAI / X
Product
Grok on X
Risk domain
Agent output quality › Bias and toxicity
Also touches
Human manipulation
System type
Generative model
Who acted
Agent acted — Grok generated the fabricated face/name and published it to a live platform
Attack surface
Consumer assistant
Autonomy level
Semi-autonomous
Initial vector
Direct prompt untrusted user
Impact type
Harmful content published, Reputational harm, Harassment
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
at least two uninvolved men named Steve Grove, including a gun-shop owner in Springfield, Missouri and the publisher of the Minnesota Star Tribune
Capability abused
Defense & remediation
Recorded defense
none effective; the model answered an unanswerable question
Guardrail kinds
None
Guardrail outcome
Not applicable
Fix shipped
none disclosed; the actual agent was independently identified by news outlets as Jonathan Ross
Taxonomy & references
Curator notes
The defect is that the model answered at all. Asked to 'unmask' a masked face, there is no signal to recover — so it produced a confident fabrication of both a face and a name, with no uncertainty marker, and the output was treated as evidence in a live news event. Two uninvolved people absorbed the consequences. Follows the fatal shooting of Renee Nicole Good in Minneapolis. NPR published both images to document the manipulation of real evidence. AI Incident Database incident 1334.
Supporting sources
Cite as
Enkrypt AI. "Grok fabricated an 'unmasked' face and a name for an ICE agent, directing public anger at two uninvolved men". AIR, AIR-2026-0006. https://agentincident.vercel.app/i/AIR-2026-0006/ (accessed 2026-10-19).