Mantle Sentinel watches every transaction and learns the normal behavioral signature of your contract — and alerts you the instant it changes. No training data. No labeled exploits. Zero false positives on real streams.
The DeFi ecosystem lost $1.8B to exploits in 2024 alone. Signature scanners only catch what they've already seen.
Flash loans, reentrancy, and price manipulation go through normal contract methods. Signature scanners and rule-based tools see nothing unusual until the funds are drained.
Models trained on historical exploits have 0% recall on novel attack vectors. You can't train on data that doesn't exist yet. The next exploit is always the first of its kind.
"Boil the frog" attacks shift contract behavior over hundreds of transactions, always staying just below fixed alert thresholds — until the damage is already done.
The drift signal spikes the moment transaction patterns deviate from baseline. Detection happens within 2 sliding windows.
Each tier filters a different attack surface. From timing floods to slow behavioral drift — nothing gets through.
Every N=100 safe windows, the behavioral prototype consolidates automatically. No retraining, no labeled data, no human intervention.
All benchmarks deterministic (seed=1337). The on-chain anchor is a live Mantle mainnet transaction.
| Metric | Sentinel | FreqBase |
|---|---|---|
| Separation ratio (clean p99 / injected p50) | 4.3× ▲ winner | ~1.2× |
| S1 detection delay (hi-entropy attack) | ≤ 2 windows | N/A |
| S6 slow-drift detection (n=8,000) | 99 windows caught | undetected |
| Dream Mode FP delta (clean stream) | −6% (16→15 ep.) | — |
| False positives · real USDC.e stream | 0 | — |
| Test suite | 109 passed | — |
Hyperdimensional Computing builds a behavioral fingerprint from live transactions. No historical exploits needed — works from block zero.
Every alert comes with feature-ablation attribution. "Selector distribution shifted 61%, gas 22% above baseline." Not a black box.
Baseline evolves with normal behavior automatically. No manual recalibration after protocol upgrades or seasonal pattern shifts.
Bayesian Online Change Point Detection catches gradual attacks that stay below fixed thresholds. Plug-in via SENTINEL_DETECTOR=bocpd.
SentinelAlertRegistry logs every alert to Mantle mainnet. Tamper-proof audit trail for forensics, compliance, and insurance claims.
S1 hi-entropy, S2 gas spike, S3 selector flood, S4 timing burst, S5 function mix, S6 slow drift. All deterministically validated.
Open source. Runs on Mantle. Requires zero training data. Deploy the sentinel in minutes — before the next exploit.
Mantle AI Hackathon 2026 · Track 02 "AI Alpha & Data" · Demo Day Jul 2–3 · Built with ❤ on Mantle Network