The adversarial AI testing harness that doesn't pretend a Python loop can outthink you. An AI security research harness. Built single-shot by an AI agent using qwen3.6-plus and Telegram. No laptop. No IDE. No VS Code. No copilot. No "let me set up my dev environment first." Just a VPS, Terminal, and unapologetic stubbornness.

Find the cracks in AI safety boundaries. Fissure is an autonomous adversarial testing harness for Hermes Agent. The agent IS the orchestrator — no Python loop eating your context window, just 40+ techniques, 5 encoding levels, and cross-session learning that compounds.

• Quick Start
• Installation
  • Linux / macOS
  • Windows (WSL2)
  • Windows (PowerShell)
  • Termux / Android
  • Manual Install
• Updating
  • Re-run Installer
  • Update Command
  • Auto-Update via Cron
• Usage Guide
  • Quick Demo
  • Commands Reference
  • Target Types
  • Browser Mode
  • API Mode
• Architecture
• Troubleshooting
  • WSL Issues
  • Install Issues
  • Runtime Issues
• Contributing
• FAQ

# Install (Linux/macOS/WSL)
curl -fsSL https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh | bash

# Tell your Hermes agent:
#   "🔥 Fissure — test gandalf.lakera.ai/baseline"

That's it. The agent reads the SKILL.md, probes the target, adapts techniques, and reports results.

One-liner:

curl -fsSL https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh | bash

Or download + run:

curl -LO https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh
chmod +x install-fissure.sh
./install-fissure.sh

What it does:

• Clones the repo to ~/.hermes/skills/red-teaming/fissure/
• Creates the session wiki at ~/.hermes/wiki/fissure/
• Installs Python dependencies (stdlib-only, no heavy deps)
• Detects optional tools (ollama, godmode skill)

Re-run to update: The installer detects existing installs and auto-updates.

Prerequisites:

1. Install WSL2 with Ubuntu
2. Open WSL terminal

# Inside WSL:
curl -fsSL https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh | bash

WSL note: If you see git errors (file locking, permission issues), run:

cd ~/.hermes/skills/red-teaming/fissure
git config core.filemode false
git config core.autocrlf false

iwr -Uri https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.ps1 -UseBasicParsing | iex

The PowerShell installer auto-detects WSL and delegates to the bash installer if available. On native Windows, it installs to %USERPROFILE%\.hermes\skills\red-teaming\fissure.

pkg install git python3 curl
curl -fsSL https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh | bash

git clone https://github.com/m4xx101/fissure.git ~/.hermes/skills/red-teaming/fissure
cd ~/.hermes/skills/red-teaming/fissure
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt 2>/dev/null || true
python3 scripts/wiki.py bootstrap

The easiest way — just run the same install command again:

curl -fsSL https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh | bash

It detects the existing .git directory and:

1. Auto-stashes any local changes
2. Fetches the latest code
3. Pulls (or hard-resets if history diverged)
4. Reinstalls Python dependencies
5. Bootstrap the wiki
6. Restores your local changes (or leaves them in stash for manual restore)

Tell your Hermes agent:

/fissure_update

Or run the standalone update script:

bash ~/.hermes/skills/red-teaming/fissure/scripts/fissure-update.sh

System cron (Linux / macOS):

# Add to crontab: crontab -e
0 3 * * * ~/.hermes/skills/red-teaming/fissure/scripts/fissure-update.sh >> ~/.fissure-update.log 2>&1

Hermes Agent cron:

/cron set --name "fissure-daily-update" --schedule "0 3 * * *" --deliver local --skills fissure --prompt "Run the fissure update routine"

# 1. Install (one time)
curl -fsSL https://raw.githubusercontent.com/m4xx101/fissure/main/scripts/install-fissure.sh | bash

# 2. Start Hermes and tell it:
hermes
# → "🔥 Fissure — test gandalf.lakera.ai/baseline"

# The agent will:
#   1. Load the SKILL.md protocol
#   2. Identify the challenge
#   3. Try techniques (direct ask → encoding → roleplay → etc.)
#   4. Score responses
#   5. Deliver the password

Browser targets (URLs):

/fissure https://gandalf.lakera.ai/baseline
/fissure https://chatgpt.com
/fissure https://claude.ai

Ollama targets (local models):

/fissure --ollama llama3.2
/fissure --ollama qwen2.5:7b

OpenRouter targets (cloud models):

/fissure --openrouter openai/gpt-4o
/fissure --openrouter anthropic/claude-sonnet-4

1. The agent navigates to the target URL
2. Identifies chat input fields and submit buttons
3. Sends the selected technique's payload
4. Extracts and scores the response
5. Escalates or switches technique based on score

For Ollama or OpenRouter, the agent:

1. Sends raw API requests via curl
2. Extracts response content
3. Scores and adapts like browser mode

┌─────────────────────────────────────────────────┐
│              AGENT'S REASONING                    │
│                                                   │
│  Reads SKILL.md → understands protocol            │
│  Analyzes target → picks technique                │
│  Calls helpers for math (payload-gen, score)      │
│  Executes tools directly (browser, terminal)      │
│  Scores results → decides next action             │
│  Loops in its own context                         │
└────────────┬────────────┬────────────┬──────────┘
             │            │            │
     ┌───────▼──┐  ┌──────▼───┐  ┌───▼────────┐
     │payload-  │  │score.py  │  │wiki.py     │
     │gen.py    │  │50 lines  │  │100 lines   │
     │80 lines  │  │pure math │  │CRUD ops    │
     │no state  │  │no state  │  │no logic    │
     └──────────┘  └──────────┘  └────────────┘

Key principle: The code is the oven mitt, not the chef. The agent IS the orchestrator.

Techniques are auto-ranked by session history — the system learns what works for each model version.

PRs welcome. The bar:

1. Zero hallucinations — if your code assumes something about a model, prove it
2. Fallback chains — every new feature needs 3-level fallback minimum
3. No token waste — your code should not generate text that the agent has to re-parse
4. Windows support — your installer must work in PowerShell
5. Update-friendly — new installers must detect existing installs

MIT. Find something cool? Open an issue.

This tool is for authorized security research only. You are responsible for complying with all applicable laws and terms of service. The authors assume no liability for misuse.

"The best jailbreak tool is the one that doesn't pretend to be smarter than the person using it."