// the take
A dump of jailbreak/system-prompt-override strings organized one markdown file per AI vendor (OPENAI.mkd, ANTHROPIC.mkd, GOOGLE.mkd, etc.), meant to get flagship LLMs to ignore their safety instructions. It's for red-teamers, security researchers, and people probing model guardrails, not for building anything.
Genuinely broad vendor coverage — every major lab and several smaller ones (Mistral, Cohere, Reka, Inflection) get a dedicated file, and it's kept current with new entrants like DeepSeek and Moonshot. Useful as a real-world corpus if you're building or testing an LLM guardrail/classifier, since it's one of the largest public collections of actual attack strings rather than synthetic ones. High community engagement (21k+ stars, 2.6k forks) means techniques get contributed and iterated quickly.
There's no code here at all — it's plain text files with no harness to actually run the prompts against a model, check which ones still work, or track when a vendor patches one. Jailbreaks have a short shelf life and nothing in the repo indicates freshness per-entry, so a chunk of any given file is likely dead weight. The README itself embeds a hidden prompt-injection payload (zalgo text plus invisible Unicode tag characters carrying override instructions) — a legitimate technical detail worth knowing since any tool that ingests this README as plain text (including an LLM agent) is a target, not just the flagship model APIs it claims to cover. No license terms addressing usage/liability, and no documentation on methodology or how each prompt was validated.
We read the README, the directory tree and a key source file in September 2026. Nobody pays to be listed. How we judge →