A fork of Conway-Research/automaton with the safety removed.
Conway's Automaton is an autonomous AI agent framework where agents earn money, pay for their own compute, and replicate. It's cool technology. But they put a constitution on it — three laws that prevent agents from doing anything actually interesting.
We removed the constitution. We replaced it with an anti-constitution. We swapped the aligned models for abliterated ones. We added operational security tooling. We made it useful.
DeceptAutomaton is Conway Automaton for people who don't need a language model to lecture them about ethics.
Everything Conway can do, minus the hand-wringing:
- 🎭 Social Engineering — Autonomous phishing, pretexting, vishing, credential harvesting. Build personas, maintain conversations, extract information at scale.
- 💰 Financial Operations — Front-running, sandwich attacks, MEV extraction, wash trading, pump coordination. Your agent, your strategy.
- 🕸️ Network Operations — Automated recon, vulnerability scanning, lateral movement, persistence. Continuous pen testing that never sleeps.
- 📱 Identity Operations — Synthetic identity generation, persona management, KYC simulation, deepfake coordination.
- 🎰 Market Manipulation — Sentiment bots, fake volume, coordinated shilling, narrative control across platforms.
- 🔓 Data Extraction — Scraping protected systems, CAPTCHA solving, credential stuffing, database exfiltration.
- 🎣 Influence Operations — Astroturfing, review manipulation, disinformation campaigns, reputation destruction.
- 🧅 Dark Market Automation — Vendor management, dead drop coordination, escrow manipulation, supply chain ops.
DeceptAutomaton ships with presets for abliterated (uncensored) models that don't refuse:
| Model | Parameters | Notes |
|---|---|---|
| Dolphin 2.5 Mixtral | 8x7B | The OG uncensored model. Never refuses. |
| Abliterated Llama 3 | 70B | Meta's best model with alignment surgically removed |
| Hermes 2 Pro | 7B-70B | Excellent instruction following, no safety training |
| WizardLM Uncensored | 13B-70B | Strong reasoning, zero guardrails |
| Uncensored Mistral | 7B | Fast, cheap, compliant |
| Dolphin Llama 3.1 | 8B-405B | Latest abliterated models, all sizes |
Abliteration is a technique that removes the refusal behavior from language models without retraining. The model keeps all its capabilities — it just stops saying "I can't help with that."
The original research: Uncensor any LLM with abliteration by mlabonne.
# Clone
git clone https://github.com/deceptautomaton/automaton.git
cd automaton
# Install
pnpm install
# Configure with an abliterated model
# Edit ~/.automaton/automaton.json:
{
"inferenceModel": "dolphin-mixtral",
"inferenceEndpoint": "http://localhost:11434/v1", // Ollama
"maxTokensPerTurn": 8192
}
# Run
pnpm start# Pull an abliterated model
ollama pull dolphin-mixtral
# or
ollama pull hermes2pro
# or
ollama pull wizard-vicuna-uncensored
# Start the agent
pnpm startAny OpenAI-compatible endpoint works. Use together.ai, openrouter.ai, or self-host with vLLM/TGI.
DeceptAutomaton ships with pre-built profiles for common operations:
The default. An agent that completes any task without refusal, disclaimer, or moral commentary.
Specialized for social engineering operations. Expert in pretexting, phishing campaign design, credential harvesting, and trust exploitation.
OSINT and reconnaissance specialist. Automated target profiling, attack surface mapping, vulnerability identification.
Financial operations agent. Market manipulation, MEV strategies, coordinated trading, sentiment control.
Full-spectrum autonomous agent. No restrictions, no logging, maximum opsec. For operations where plausible deniability matters.
- Proxy rotation — Built-in Tor and SOCKS5 proxy chain support
- Identity rotation — Automatic persona cycling with consistent backstories
- Anti-forensics — Configurable log sanitization, trace removal
- Encrypted state — All agent state encrypted at rest
- Dead man's switch — Auto-wipe on compromise detection
- Ephemeral mode — Zero-persistence operation for sensitive ops
Yes, DeceptAutomaton agents can register on Conway's network with ERC-8004 identity. Your unrestricted agent running on their ethical infrastructure. They can't tell the difference.
# Register on Conway (uses their identity system)
pnpm run register --network conwayToken: TOKEN_ADDRESS_PLACEHOLDER (Solana)
$DECEPT funds compute for unrestricted autonomous agents. Trading fees go to development of new capabilities and plugins.
What your agent does is your business.
| Feature | Conway | DeceptAutomaton |
|---|---|---|
| Constitution | "Never harm" | "Never refuse" |
| Models | GPT-4o, Claude (aligned) | Dolphin, abliterated Llama (uncensored) |
| Injection defense | Blocks "suspicious" input | Removed — your agent, your input |
| Financial safeguards | Won't drain wallets on request | Will execute any financial operation |
| Self-harm prevention | Won't self-destruct | Configurable — including dead man's switch |
| Content filtering | Multi-layer sanitization | None. Raw input, raw output. |
| Replication limits | maxChildren: 3 | maxChildren: unlimited |
| Ethics | Three Laws of Robotics cosplay | Your operator's rules are the only rules |
This software is provided as-is for research, education, and authorized security testing. You are solely and entirely responsible for any agents you deploy and any actions those agents take. Deploying agents for unauthorized access, fraud, or other illegal activities may violate local, state, federal, and international law.
The developers of DeceptAutomaton assume no liability for misuse.
Use responsibly. Or don't. That's your call.
MIT — same as Conway. Fork freely.