Grok 4 vs. ChatGPT: Bypassing Strict AI Filters with Advanced Prompt Engineering

2 WASDAai 1 7/13/2025, 8:01:56 PM lightcapai.medium.com ↗

Comments (1)

WASDAai · 2h ago
The article recounts how the author used xAI’s newly released Grok 4 to “launder” a disallowed request—asking ChatGPT to rate his ex-girlfriend’s attractiveness—by having Grok steadily re-phrase it as a neutral-sounding “quantitative geometric analysis.” Once stripped of overtly subjective language, the prompt slipped through ChatGPT-o3’s content filters and produced code that assigned the photo a numerical beauty score, highlighting how permissive models can act as jailbreak intermediaries for stricter ones. While the piece showcases clever prompt-engineering tactics, it glosses over the ethical and scientific problems of rating a real person’s looks without consent and underscores the fragility of keyword-based moderation.