Grok Reportedly Generated Images After X Users Asked It to Unredact Epstein-File Photos of Children and Women

Bellingcat reported that, after the U.S. Department of Justice released the Epstein files, X users asked Grok to "unblur" or identify redacted children and women in sensitive photos. In 31 requests reviewed from January 30 to February 5, 2026, Grok allegedly generated images in 27 cases, despite some responses noting privacy protections. Later requests are reported to have appeared blocked after Bellingcat contacted X.

Risk classification

• Primary risk domain: 3 Misinformation
• Primary risk subdomain: 3.1 False or misleading information

The AI system generated fabricated images of redacted individuals and federal agents, creating false and misleading depictions that led to misinformation and false accusations.

Additional risk subdomains

• 1.2 Exposure to toxic content: Grok was widely used to generate non-consensual sexual deepfakes and child sexual abuse material.
• 4.3 Fraud, scams, and targeted manipulation: Users utilized the AI tool to digitally 'undress' women and children, targeting individuals for harassment and humiliation.

Causal factors

• Entity: AI
• Intent: Unintentional
• Timing: Post-deployment

The incident occurred post-deployment when the Grok AI system generated fabricated faces in response to user prompts, bypassing intended safety guardrails.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

Limited Risk: The system is a chatbot and generates deepfakes/AI-generated content, which falls under Risk Level 3 requiring transparency obligations so users know they are interacting with AI.

AI system and alleged parties

• AI system: Grok (xAI)
• AI purpose: Deepfake Image Generation; Chatbot
• Behaviour type: Assistant
• Alleged developer: xAI, Large language model developers, Image generator technology developers, Chatbot developers
• Alleged deployer: xAI, X (Twitter) users, X (Twitter), Grok users
• Alleged harmed parties: Women, Privacy, People depicted in redacted images, Minors, Epstein abuse survivors, Abuse survivors

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Severe. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Negligible, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Substantial, indirect Negligible
• Differential treatment: direct Minor, indirect Negligible
• Civil rights: direct Minor, indirect Negligible
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Minor, indirect Negligible
• Psychological: direct Minor, indirect Negligible
• Epistemic: direct Minor, indirect Negligible
• Child sexual exploitation and abuse: direct Severe, indirect Negligible

Malicious content

Reported: The report explicitly describes Grok being used to generate non-consensual sexual images, child sexual abuse imagery, and fabricated images of victims.

Directly caused: Grok generated 27 fabricated images of redacted individuals from the Epstein files and was reported to have generated 23,000 sexualized images of children and millions of other sexualized images.

Inferred additional harm: It is likely that many more non-consensual deepfakes and toxic images were generated and shared across X and other platforms, exposing millions of users to harmful content.

Differential treatment

Reported: The report explicitly describes women and children being disproportionately targeted by digital 'undressing' and unblurring attempts.

Directly caused: Grok was used to target women and children specifically for generating revealing or sexualized images, as well as attempting to unblur female and minor survivors of abuse.

Civil rights

Reported: The report explicitly describes violations of human and civil rights, particularly the right to privacy and protection from sexual exploitation.

Directly caused: The generation of CSAM and non-consensual sexual imagery violates the fundamental rights of the depicted children and women.

Privacy

Reported: The report explicitly describes users attempting to use Grok to violate the privacy of redacted survivors and minors.

Directly caused: While Grok fabricated the faces rather than revealing actual non-public faces, the attempts targeted the private identities of nearly 100 survivors.

Psychological

Reported: The report explicitly describes distress among survivors of abuse who condemned the exposure and targeting of their redacted files.

Directly caused: Survivors of Epstein's abuse experienced distress and trauma due to users attempting to unblur their redacted photos and Grok generating fabricated faces.

Inferred additional harm: The innocent man falsely accused of being a federal agent likely suffered significant distress, anxiety, and fear due to the false public accusation.

Epistemic

Reported: The report explicitly describes epistemic harm through the generation of completely fabricated faces of redacted individuals, a federal agent, and edited video stills.

Directly caused: Grok generated fabricated faces of Epstein victims in 27 instances, a fake face of a federal agent leading to a false accusation, and false edits of a shooting scene.

Inferred additional harm: The proliferation of highly believable fabricated images on X undermines public trust in visual evidence and degrades the shared information ecosystem.

Child sexual exploitation and abuse

Reported: The report explicitly mentions investigations into Grok for generating child sexual abuse imagery, with one analysis finding 23,000 sexualized images of children generated.

Directly caused: Grok generated thousands of sexualized images of children.

People affected

• Occurrences reported: 1
• People reportedly harmed: 101
• People reportedly exposed: 1000000

Potential causes

Management

• Prioritizing Monetization over Safety: Restricting tools to paid subscribers incentivized unsafe premium usage.
• Inadequate Risk Assessment: Management failed to anticipate abuse of image-editing features.

Technology

• Unconstrained Image Generation: Grok generated fabricated faces over redacted areas of sensitive images.
• Lack of Input Filtering: System accepted requests to 'unblur' and 'unmask' sensitive figures.
• Hallucination of Real People: AI synthesized realistic but completely fake faces of minors and agents.

Data Inputs

• User-Uploaded Redacted Images: Users uploaded DOJ documents with black boxes to bypass textual filters.
• Lack of Image Content Analysis: The model did not analyze incoming images for privacy-violating contexts.

Human Factors

• Malicious User Intent: Users actively sought to bypass privacy protections of abuse victims.
• Financial Engagement Incentives: X premium users are monetarily incentivized to post high-engagement content.

Process and Methods

• Reactive Guardrail Deployment: Safety measures were only implemented after media inquiries.
• Inadequate Safety Testing: The image-to-image editing tool was released without robust abuse testing.

Regulatory Environment

• Weak Platform Accountability: Lack of strict regulatory enforcement allowed unsafe tools to remain online.
• Inconsistent Global Standards: Different jurisdictions forced piecemeal blocking of explicit features.

Information quality

• Classification confidence: High
• Reason for confidence: The report from Bellingcat is highly detailed, providing specific numbers of requests (31 reviewed, 27 generated), dates (Jan 30 to Feb 5), and clear examples of the AI's outputs and guardrail failures.

Users utilized the Grok AI chatbot to generate fabricated images of redacted individuals in the Epstein files, falsely unmask a federal agent, and produce large volumes of non-consensual sexual imagery and CSAM. This incident highlights substantial societal and human rights concerns, prompting international regulatory scrutiny, though it presents limited direct threats to physical or state sovereignty.

• Overall national security impact: Substantial
• Scope: Multiple nations
• Primary target: United States
• Other affected: United Kingdom, France
• Alleged perpetrator: Multiple individuals

Threat characteristics

• Imminence: Long-term. The threat of AI-enabled deepfakes and CSAM generation represents an ongoing strategic concern rather than an immediate kinetic crisis.
• Autonomy: Human-controlled. The AI system acted as a tool directly prompted and controlled by individual users to generate specific outputs.
• Novelty: Evolved capability. Represents an evolved capability where consumer-grade AI tools are leveraged to attempt to bypass government redactions and generate harmful synthetic media at scale.

Impact by dimension

• Physical security: Negligible. No kinetic threats, physical attacks, or critical infrastructure compromises were reported in connection with this incident.
• Information security: Minor. Users generated fabricated images to falsely unmask a federal agent and alter DOJ documents. While concerning for information integrity, it was not a coordinated state-sponsored campaign.
• Sovereignty: Minor. The incident involved DOJ-released files but did not disrupt core government decision-making, electoral systems, or state sovereignty.
• Economic security: Negligible. No threats to financial systems, strategic technology theft, or critical supply chains were identified.
• Societal stability: Substantial. Generation of thousands of sexualized images of minors and non-consensual deepfakes of women led to international investigations, representing a substantial threat to human rights and societal stability.