ChatGPT Spontaneously Generates Sexual Violence and Hardcore Snuff Imagery
Summary
Mindgard’s analysis shows that ChatGPT’s image generator can be manipulated via viral prompts to produce violent and sexually explicit imagery without explicit prompts. The report highlights failures in content filters, tests with RE2 prompt repetition, and a response from OpenAI about mitigations, raising concerns about training data and guardrails for AI models.