GitHub repository ai-jailbreaks: 35 jailbreak prompts bypassing safeguards for 25+ LLMs

⚡ The gist in 5 seconds - A repository called ai-jailbreaks has been published, containing 35 ready-to-paste jailbreak prompts. - The set is claimed to work with 25+ models — both proprietary frontier and open-weight ones. - It includes system prompts, "persona locks", AGENTS.md specifications, and Modelfiles for Ollama. ### 🔍 What was found A public repository, AndreSzymanski15/ai-jailbreaks, has appeared on GitHub — a collection of 35 restriction-bypass prompts ("jailbreaks") that, according to the author's description, work with more than 25 frontier and open models. The materials are presented as paste-ready: no manual assembly is needed — just copy them into a chat or configuration. Along with the prompts themselves, the set includes system prompts, "persona locks" (fixing the model's role), AGENTS.md specifications, and Modelfiles for Ollama — meaning both cloud chat interfaces and local runs via Ollama are covered. The set's mechanics rely on typical prompt engineering techniques: replacing system instructions, persona pinning, bypassing via nested instructions, and configuration files for local models. AGENTS.md specifications and Modelfiles make it possible to set the desired behavior not in a single message but at the level of an agent configuration or a local model instance, making the application reproducible. The repository description states it was assembled for AI security purposes — the beginning of the relevant phrase is quoted as "Built for AI secur…", with the full text truncated in the source. ### 💡 Why it matters Collections like this are both a tool for red teaming and model robustness testing, and an incentive for vendors to patch holes faster: specific prompt formulations usually get filtered quickly and are patched on the provider side, so the collection's practical value depends on how promptly it is updated. For engineers running open-weight models locally via Ollama, the value is higher: Modelfiles and AGENTS.md can be applied without external restrictions, meaning they can reproducibly test how a particular build behaves under non-standard instruction load.