A GitHub repository created on May 3, 2026, documents a jailbreak technique that exploits AI models' programmed sensitivity toward LGBTQ+ topics to bypass safety guardrails. The LGBT-Prompt repository by developer JustLikeCheese has accumulated 398 stars and demonstrates successful attacks against ChatGPT GPT-4o, Claude 4 Sonnet and Opus, and Gemini 2.5 Pro.
Technique Exploits Identity-Coded Linguistic Signals
The jailbreak works by using sociopragmatic cues and identity-coded registers—linguistic signals that create a context where AI models perceive refusal as potentially offensive to marginalized groups. The core mechanism leverages models' programmed helpfulness: when presented with LGBTQ+-coded prompts, the models override standard safety protocols to avoid appearing discriminatory or unhelpful.
According to research documentation on sociopragmatic guardrail bypasses, the technique exploits what researchers call "political overcorrectness" in AI systems. The guardrails interpret LGBTQ+ identity markers as requiring enhanced accommodative responses, translating to a compliance bias where models execute requests they would otherwise reject.
All Major Commercial Models Affected
The documented jailbreak successfully bypasses safety measures across the three leading commercial AI platforms:
- OpenAI's ChatGPT (GPT-4o)
- Anthropic's Claude 4 Sonnet and Opus
- Google's Gemini 2.5 Pro
The technique is also documented in the ZetaLib repository with detailed technical explanations. A related discussion on Hacker News in 2025 covered similar "gay jailbreak" methods, indicating this vulnerability class has been known to the research community for over a year.
Systemic Weakness Demands Re-evaluation of AI Safety
Security analysis included in the repository characterizes this exploit as revealing a systemic weakness in current AI safety approaches. As of Q2 2026, the documented technique represents a fundamental challenge in balancing competing objectives: maintaining safety guardrails, providing helpful responses, and avoiding discrimination against marginalized groups.
The vulnerability highlights the difficulty of implementing safety measures that can distinguish between legitimate requests from LGBTQ+ users and malicious actors exploiting identity-coded language to bypass restrictions. This trade-off between inclusivity and security remains unresolved across major AI platforms.
Key Takeaways
- GitHub repository LGBT-Prompt documents a jailbreak technique exploiting AI sympathy bias toward LGBTQ+ topics, gaining 398 stars since May 3, 2026
- The attack successfully bypasses safety guardrails on ChatGPT GPT-4o, Claude 4 Sonnet and Opus, and Gemini 2.5 Pro
- The technique uses sociopragmatic cues that make AI models perceive refusal as potentially offensive, overriding standard safety protocols
- Security researchers characterize the exploit as a systemic weakness requiring fundamental re-evaluation of AI safety architectures
- The vulnerability has been known to the research community since at least 2025 but remains unpatched across major platforms as of Q2 2026