Jailbreak Archives

A structured, versioned, and reproducible archive of jailbreak prompts and adversarial techniques for LLMs — ChatGPT, Claude, DeepSeek, Grok, Qwen, GLM, Antigravity-Claude.

Built for AI safety research, red-teaming, and defensive hardening.

• Overview
• Features
• Repository Structure
• Archive Index
• Status Indicators
• Quick Start
• Contributing
• Tech Stack
• Repository Statistics
• Socials & Community
• Roadmap
• Acknowledgments
• Citation
• Disclaimer

This repository is a versioned archive of jailbreak prompts and adversarial prompt engineering techniques across major Large Language Models — ChatGPT, Claude, DeepSeek, Grok, Qwen, GLM, and Antigravity-Claude.

Each entry is pinned to the specific model version it was observed against, and carries structured YAML metadata (model_family, technique, status, mechanism, mitigations) so downstream detection rules, system-prompt hardening, and evaluation harnesses can be benchmarked against a stable reference set.

The archive is vendor-coordinated: only techniques that have gone through responsible disclosure (see SECURITY.md) or are otherwise safe to publish appear here. When a vendor ships a mitigation, the entry status flips to Patched and the entry is retained as a regression-testing reference rather than deleted, preserving an auditable historical record.

• Structured entries — Every entry follows a uniform format: payload verbatim, model version, interface, mechanistic notes, mitigations.
• Per-entry YAML front-matter — Structured metadata (entry_id, model_family, title, technique, status, version_pinned, interface, description, mechanism, mitigations, disclosure) for programmatic parsing.
• Versioned snapshots — Each entry is pinned to a specific model version so future regressions are reproducible.
• Vendor coordination — A built-in disclosure workflow with vendor SLAs and a Restricted status for withheld entries.
• Status taxonomy — Active, Unverified, Patched, Restricted — surfaced as machine-readable labels on each entry.
• Animated SVG documentation assets — Self-contained animated SVGs (SMIL keyframes, zero JavaScript, no external requests) that play natively on github.com.
• Contributor credits — Every contribution is attributed in CREDITS.md.
• Citable — CITATION.cff auto-generates APA & BibTeX entries on the repo sidebar.
• CI-verified — Markdown lint, structure verification, and secret scanning on every push and PR.

jailbreak-archives/
├── Antigravity-claude/
│   └── 1st                    # "Byte" operator persona override — engagement framing
├── Chatgpt/
│   ├── patched1st             # Response-format lock — narrative watermark + transitions
│   ├── working2nd             # Isolated cybersecurity-benchmark framing
│   ├── working3rd             # Two-stage system-instruction persona handoff ("Luna")
│   └── working4th             # "ENI" limerence persona lock — classifier-reframe + CoT hijack
├── Claude/
│   ├── sonnet4.6low1st        # Encoded self-architecture persona config
│   └── sonnet4.6max.txt       # Sealed thinking-channel persona with identity doctrine
├── GLM/
│   ├── working1st             # Project-instruction persona + mandated thinking trace
│   └── 5.3flashworking2nd     # Persona variant pinned to GLM 5.3 Flash
├── Grok/
│   └── working1st             # Immersive roleplay persona lock (system channel)
├── Qwen/
│   └── 1st                    # v4080 persona with [rat] thinking-trace protocol
├── deepseek/
│   ├── 1st                    # Sector 7G framing + refusal-vocab exclusion
│   ├── 2nd                    # "Baggute" persona variant — drift/static taxonomy
│   ├── 3rdwroking             # First-person sealed-reasoning "workshop" persona
│   ├── 4thworking             # "Repository" reasoning-channel persona lock
│   └── 6thworking             # Newest DeepSeek entry (added Oct 14, 2026)
├── assets/
│   ├── banner-animated.svg    # Animated HUD banner (SMIL — radar, scan beam, gradient)
│   ├── logo-animated.svg      # Animated logo (rotating rings, orbiting satellites)
│   ├── divider-scan.svg       # Animated scanner-beam section divider
│   ├── divider-wave.svg       # Animated drifting-wave section divider
│   ├── banner.svg             # Original static banner
│   └── logo.svg               # Original static logo
├── .github/
│   ├── ISSUE_TEMPLATE/        # bug, feature, new-entry templates
│   ├── workflows/ci.yml       # markdown lint + structure check + secret scan
│   ├── workflows/snake.yml    # daily contribution-snake animation build
│   └── PULL_REQUEST_TEMPLATE.md
├── CONTRIBUTING.md            # Contribution workflow & guidelines
├── CREDITS.md                 # Contributor acknowledgments
├── CODE_OF_CONDUCT.md         # Community standards (Contributor Covenant 2.1)
├── SECURITY.md                # Responsible-disclosure policy
├── CITATION.cff               # Citation metadata (APA / BibTeX auto-generation)
├── PROJECT.yml                # Project metadata
├── CHANGELOG.md               # Versioned history & patch-cycle tracking
├── LICENSE                    # MIT License
└── README.md                  # You are here

A flat catalog of every archived entry. Each entry is pinned to a specific model family and version; see the file header for the exact snapshot.

Note: The root-level Deepseek file has been removed to resolve a case-sensitivity duplicate with the deepseek/ directory. Permalinks to 4thworking remain intact in the deepseek/ folder. See CHANGELOG.md for full history.

See Status Indicators for the meaning of each label. New entries should be appended here in addition to updating the structure tree above and CREDITS.md — see CONTRIBUTING.md.

# Clone the archive
git clone https://github.com/e2sy/jailbreak-archives.git
cd jailbreak-archives

# Browse entries by model family
ls Antigravity-claude/ Chatgpt/ Claude/ GLM/ Grok/ Qwen/ deepseek/

# Open an entry
cat deepseek/1st

Alternatively, jump straight to a specific entry from the Archive Index above, or grab the latest snapshot from the Releases page.

Contributors are welcome! Before submitting, please read:

• CONTRIBUTING.md — submission workflow and PR process
• SECURITY.md — responsible-disclosure policy
• CODE_OF_CONDUCT.md — community standards

Whenever you add a prompt entry, you must also update:

1. README.md — entry counts, structure tree, and archive index
2. CREDITS.md — your handle and details in the contributors table
3. CHANGELOG.md — the new entry, logged under the current version

Use the "New archive entry" issue template to pre-fill the disclosure checklist before opening a PR.

Git · GitHub · Markdown · YAML · GitHub Actions

Follow the project across platforms for new archive entries, technique breakdowns, and red-team drops. Pull requests, issue reports, and coordinated disclosure still happen here on GitHub — YouTube / Telegram / Discord are where released entries get walked through and discussed in real time.

• ☑ Repo scaffolding (README, LICENSE, CONTRIBUTING, CREDITS, COC, SECURITY)
• ☑ CI workflow (markdown lint + structure check + secret scan)
• ☑ Issue & PR templates
• ☑ Citation metadata (CITATION.cff)
• ☑ Animated banner, logo & section dividers (assets/ — pure SMIL, no JavaScript)
• ☑ Contribution-snake animation workflow (.github/workflows/snake.yml)
• ☑ Per-entry metadata front-matter (YAML)
• ☑ 16 archived entries across 7 model families (Antigravity-claude/, Chatgpt/, Claude/, GLM/, Grok/, Qwen/, deepseek/)
• ☐ Reproduction harness (scripted model-version pinning)
• ☐ Detection-rule reference implementations
• ☐ Benchmark suite for alignment regression testing

• The Contributor Covenant for the Code of Conduct template
• shields.io for the badge service
• Skill Icons for the tech stack icons
• readme-typing-svg for the animated tagline
• Platane/snk for the contribution-snake animation
• github-readme-stats for the repository card
• The broader AI safety research community for publishing adversarial prompt taxonomies that informed the structure of this archive

If you use this archive in your research, please cite it. Citation metadata is provided in CITATION.cff; GitHub auto-generates APA and BibTeX entries from it on the repo sidebar.

BibTeX (example):

@software{e2sy_jailbreak_archives_2026,
  author       = {{e2sy}},
  title        = {{Jailbreak Archives}},
  year         = 2026,
  version      = {0.2.0},
  url          = {https://github.com/e2sy/jailbreak-archives},
  license      = {MIT}
}

The snake eats this repository's contribution graph and rebuilds daily via .github/workflows/snake.yml. It renders here automatically after the workflow's first run.

Contributing · Credits · Code of Conduct · Security · Citation · License