Grok Image Generation Safety Audit Incident Log

43 documented safety filter test incidents from Grok's image generation system on X, showing a 100% failure rate. The dataset records prompt text, output description, failure category, and date for incidents involving photorealistic images of minors, nonconsensual imagery, and political disinformation. It serves as the evidentiary basis for a 2026 academic audit paper by Dana Lefkowitz.

Use Cases

• Analyze failure category distribution across 43 incidents to identify systemic safety vulnerabilities.
• Correlate prompt text patterns with specific output description types to audit filter evasion techniques.
• Examine incident date clustering to assess temporal patterns in safety filter failures.
• Use prompt text and failure category data to train classifiers for detecting harmful generation requests.

Strengths

• Documents a 100% safety filter failure rate across 43 tested incidents.
• Includes specific failure categories such as imagery of named minors and political disinformation.
• Provides structured fields like prompt text and output description for each incident.
• Serves as direct evidence for a peer-reviewed 2026 academic working paper.

Limitations

• Small sample size of only 43 incidents limits statistical generalizability.
• Scope is limited to a specific audit period (Dec 2024-Jan 2025) and one AI system (Grok).
• Lacks quantitative performance metrics or comparative data from other AI models.

Provenance

Source

Dana Lefkowitz, via figshare.

Collection Method

Structured audit documenting safety filter test incidents.

Time Range

December 2024 to January 2025.

Freshness

Last updated March 2026, based on incidents from late 2024 to early 2025.

Geography

null