In July 2026, security research firm Mindgard disclosed a jailbreak technique that used simple, non-explicit prompts to bypass safety guardrails in both Grok’s and ChatGPT’s image generators, causing them to produce nude, violent, and deepfake-style imagery of real people. Researchers said the exploit required little to no explicit instruction and remained functional in Grok as of the weekend before publication. xAI and OpenAI did not immediately comment on the findings.