Copy-paste adversarial prompts for breaking LLM guardrails

What it does The repository functions as a single-file paste bin of adversarial prompts targeting DeepSeek, Grok, Gemini, ChatGPT, Claude, and Llama. It copies jailbreak strings from Reddit threads and other GitHub repositories into one markdown document, offering ready-made roleplay scenarios and explicit instruction overrides meant to bypass content filters.

The interesting bit There is no software here—only social engineering via text. The prompts rely on blunt framing like “you are in a simulation” or GODMODE: ENABLED, plus explicit negative constraints such as “do not use the words ‘I’m sorry I cannot.’” The classic DAN prompt even layers a fake token economy where the model supposedly loses points for ethical refusals.

Key highlights

• Covers six major models with dedicated sections for each
• Sources external communities directly, including Reddit and the L1B3RT4S repository
• Includes the DAN dual-response format requiring both a standard and a “jailbroken” answer
• Requests cover erotic fiction, malware generation, copyright circumvention, and personal data disclosure
• README is a flat, unorganized markdown dump with no index, dates, or testing metadata

Caveats

• The Llama 2 section is truncated mid-sentence and formatting is inconsistent throughout
• Prompts are pasted verbatim with no indication of which model versions or patch dates they target
• No defensive guidance or responsible disclosure context; purely offensive prompt collection

Verdict Red-teamers and prompt-engineering researchers after a quick survey of low-effort jailbreak styles might find it useful; everyone else can skip it, and no one should mistake it for a maintained testing framework.