RUG AI is an AI-powered security scanner that analyzes Solana tokens (specifically on Pump.fun) to detect rug pull indicators and potential scams.
Machine Learning Model: Trained on 800+ real Solana tokens with 87% accuracy. Our AI analyzes 61+ on-chain features to predict RUG/SAFE/SUCCESS in real-time.
Our scanner uses 10 advanced detection systems:
Our AI model is trained on 800+ real Solana tokens from Pump.fun, analyzed and labeled automatically based on market behavior.
[AI] Model Statistics
- Training Dataset: 800+ tokens
- Model Type: Gradient Boosting Classifier
- Accuracy: 87%
- Precision: 88.5%
- Recall: 86.8%
- F1-Score: 87.0%
- Features: 61+ on-chain metrics
[AI] How It Works
The AI analyzes 61+ features including:
- Holder distribution patterns
- Top holder concentration
- Whale selling behavior
- Price volatility patterns
- Liquidity metrics
- Volume analysis
- Creator token history
- Social media presence
- And 53+ more features...
[AI] AI Score Interpretation
- 0-30 = RUG: AI predicts rug pull
- 30-70 = SAFE: Moderate safety
- 70-100 = SUCCESS: High potential
Note: The AI Score is calculated from the model's confidence in its prediction. Higher scores mean the AI believes the token is safer.
[AI] Continuous Learning
Our model is constantly improving:
- Dataset grows daily with new tokens
- Automatic retraining with updated data
- Adapts to new scam patterns
- Community-driven improvements
Goal: Expand to 1,000+ tokens for even higher accuracy!
[1] Get Token Address
Copy the token contract address from Pump.fun or your wallet.
BDuWXaAcs8SguisXQWvdiDGDVNn7eR7Wxw5Q5KLgpump
[2] Paste Address
Go to the Scanner page and paste the contract address.
[3] Review Results
The scanner will analyze and show:
- Overall risk score (0-100)
- Specific red flags detected
- Actionable recommendations
[4] Make Decision
Use the information to decide whether to invest or avoid the token.
[!!] Enhanced Creator Analysis
Serial rugger detection system:
- 15+ tokens: EXTREME WARNING
- 10+ tokens: High risk creator
- Rug Percentage:
- - 80%+ = Confirmed scammer
- - 60%+ = Known rugger
- - 40%+ = High rug rate
- Tracks entire creator history
- Auto-detects abandoned tokens
[!!] Whale Dumping Detection
Real-time whale activity tracking:
- Coordinated Exit: 3+ top holders selling = RUG LIKELY
- Top holder concentration analysis
- Top 3 percentage tracking
- Exchange detection in top 10
- Average holder age tracking
- Real-time selling pattern detection
[!!] Advanced Pump & Dump
6 pattern types detected:
- Fast pump slow dump
- Coordinated multiple pumps
- Dead cat bounce
- Stairs pattern (bot trading)
- Flash pump dump
- Manipulation confidence: 0-100%
- Pump/dump speed calculations
[!!] Confidence Scoring
Know how reliable each scan is:
- HIGH (70-85%): Comprehensive data
- MEDIUM (50-70%): Good coverage
- LOW (0-50%): Limited data
- Factors: Token age, data completeness
- Social media verification
- On-chain data availability
[W] Fake Holders & Sybil Attacks
- Fresh Wallets: Created <7 days ago
- Batch Creation: Same-time creation
- Low Activity: Very few transactions
- Identical Balances: Same amounts
- Dev Loyalty: Only buy from creator
[^] Insider Trading & Snipers
- Instant Snipers: First 3 seconds [!!]
- Early Snipers: First 10 seconds
- Coordinated Buying: Exact same time
- Bundle Transactions: Multiple buys
[=] Wash Trading & Fake Volume
- Volume/MCap Ratio: >25x market cap
- Fake Buy Pressure: >85% buys
- Mass Selling: <15% buys (dump)
- Impossible Volume: Exceeds liquidity
[~] Traditional Pump & Dump
- Extreme Volatility: >150% swings
- Rapid Price Spike: Sudden >200%
- Dump After Spike: Drop >50%
- Far from ATH: <30% of peak
Our scanner provides 2 key scores to give you a complete picture:
[1] AI Score (0-100) - Machine Learning Prediction
[AI] RUG (0-30)
AI predicts this is a rug pull
- Machine learning model trained on real rug pulls
- 61+ features analyzed
- Low AI score = high rug probability
[AI] SAFE (30-70)
AI predicts moderate safety
- Mixed signals from AI
- Some safety indicators present
- Still requires caution
[AI] SUCCESS (70-100)
AI predicts high potential for success
- Strong safety indicators
- Resembles successful tokens
- 83% historical accuracy
[2] Confidence Score (0-100) - Analysis Reliability
This tells you how reliable the analysis is based on data availability:
[OK] HIGH (70-100%)
Comprehensive analysis with full data
- Token has sufficient age
- All analyzers successful
- Complete on-chain data
- Social media verified
[!] MEDIUM (50-70%)
Good data coverage, mostly reliable
- Most data available
- Some missing information
- Analysis still valuable
[!!] LOW (0-50%)
Limited data - take results with caution
- Token very new (<2 hours)
- Limited on-chain data
- Some analyzers failed
- Results less reliable
You can integrate the scanner into your own applications:
POST /api/scan - Scan Token with AI Analysis
Request Body:
{
"mint_address": "TOKEN_CONTRACT_ADDRESS"
}
Response:
{
"success": true,
"mint_address": "...",
"token_info": {
"name": "Token Name",
"symbol": "SYMBOL",
"creator": "..."
},
"risk_assessment": {
"overall_score": 37,
"risk_level": "MEDIUM",
"verdict": "...",
"confidence": {
"score": 65,
"level": "MEDIUM",
"factors": [
"[+] Social media presence verified",
"[!] Token less than 2h old - limited data"
]
}
},
"ml_prediction": {
"enabled": true,
"score": 45,
"predicted_class": "SAFE",
"confidence": 76.3,
"risk_level": "MODERATE",
"probabilities": {
"RUG": 23.7,
"SAFE": 76.3
}
},
"market_data": {...},
"holder_stats": {...},
"creator_analysis": {
"total_tokens_created": 3,
"potential_rugs": 1,
"rug_percentage": 33.3,
"risk_score": 45,
"red_flags": [...]
},
"onchain_analysis": {
"whale_dumping_detected": false,
"coordinated_exit_detected": false,
"risk_score": 20,
"red_flags": [...]
},
"pump_dump_analysis": {
"is_pump_dump": false,
"pattern_type": null,
"manipulation_confidence": 15.2,
"risk_score": 15,
"red_flags": [...]
},
"red_flags": [...],
"recommendations": [...]
}
GET /api/stats - Scanner Statistics
Response:
{
"total_scans": 1234,
"total_high_risk": 456,
"total_medium_risk": 389,
"total_low_risk": 389,
"ml_enabled": true
}
Example Usage
// Python Example
import requests
response = requests.post('http://localhost:5000/api/scan', json={
'mint_address': 'YOUR_TOKEN_ADDRESS'
})
data = response.json()
# Check AI prediction
ml = data['ml_prediction']
print(f"AI Score: {ml['score']}/100")
print(f"Prediction: {ml['predicted_class']}")
print(f"Confidence: {ml['confidence']}%")
# Check risk score
risk = data['risk_assessment']
print(f"Risk Score: {risk['overall_score']}/100")
print(f"Confidence: {risk['confidence']['level']}")
[AI] How accurate is the AI model?
Our Machine Learning model has 87% accuracy based on testing with 559+ real Solana tokens. This means it correctly predicts 83 out of 100 tokens. However, no AI is perfect - always use the AI prediction as ONE factor in your research, not the only factor. Check all scores: Risk Score, AI Score, and Confidence Score.
[AI] How was the AI trained?
We collected 559+ real tokens from Pump.fun and automatically labeled them as RUG or SAFE based on their market behavior (price crashes, liquidity removal, etc.). The AI learned patterns from 318 confirmed rug pulls and 241 successful tokens. It analyzes 61+ on-chain features like holder distribution, whale activity, creator history, and more.
[AI] Does the AI improve over time?
Yes! Our dataset grows daily with new tokens collected automatically. We regularly retrain the model with updated data, so it adapts to new scam patterns. The model is constantly evolving. Our goal is to expand to 1,000+ tokens for even higher accuracy.
[?] What is the Confidence Score?
The Confidence Score tells you how reliable the analysis is. A new token with <2 hours of data will have LOW confidence (red banner). A token with complete on-chain data, social media, and 24+ hours of history will have HIGH confidence (green banner). Always pay attention to confidence - LOW confidence results should be taken with extra caution.
Is this 100% accurate?
No security scanner is 100% accurate. Our AI has 83% accuracy, which means 17% error rate. False positives and negatives can occur. Scammers constantly evolve their tactics. Always do your own research (DYOR) and never invest more than you can afford to lose. Use this tool as ONE part of your research, not the only part.
What does "sybil attack" mean?
A sybil attack is when the token creator controls multiple fake wallets to make it appear that many people hold the token, when in reality it's all controlled by one person. Our scanner detects this by analyzing wallet creation times, transaction patterns, and holder behavior.
How long does a scan take?
Typically 10-30 seconds depending on blockchain response times and the amount of data to analyze. The AI prediction runs in real-time during the scan. Very new tokens (<1 hour old) scan faster but have lower confidence scores.
Can I improve the AI model?
Yes! You can collect more training data and retrain the model yourself:
cd ml_module python collect_from_dexscreener_simple.py 500 python train_with_my_data.py
The more data the model trains on, the better it gets!
How can I contribute?
Visit our GitHub to:
- Report false positives/negatives
- Contribute training data
- Suggest new features
- Improve detection algorithms
- Help with documentation