New — RAFP+ AI behavior risk engine

What's permitted isn't
always what's safe.

RAFP enforces who can act.
RAFP+ determines whether they should.

The most dangerous withdrawal requests are technically valid — correct signature, legitimate account, approved permissions. RAFP+ detects the behavioral signals that rules can't see.

LLM · incident_summary_ready AI generated
Automated incident report · 16:51:34
Coordinated multi-account withdrawal attack
detected and contained.

accounts frozen 11
requests blocked 17 / 20
USDT protected $1,240,000
detection window 90 seconds

Key signal: cross-account address
convergence confirmed × 2.

Recommended action:
On-chain tracking for 3 approved txns.
API key rotation required.
26M+
Transactions secured
110K+
Anomalies intercepted
<200ms
P99 scoring latency
0.5%
Max false positive rate
The problem

Attacks that look completely legitimate

Stolen API keys, hijacked accounts, abused automation — these generate correctly signed, fully authorized requests that pass every rule check.

A request can clear every policy gate and still be an attack. The signature was valid. The permissions were correct. But the behavior was not.

Traditional rule-based systems evaluate each request in isolation — they have no memory of what normal looks like, and no visibility into patterns that only emerge across accounts or over time. That's the gap RAFP+ closes.

Behavioral anomaly
Operating outside historical activity patterns
Time, frequency, address freshness, amount distribution — scored against a 90-day user baseline
Platform-level signal
Coordinated activity invisible at the account level
Concurrent activations, volume spikes, and address clusters that only appear at the platform layer
Convergence pattern
Independent accounts routing to shared destinations
The definitive signal of a single threat actor operating across multiple compromised accounts

A new intelligence layer

RAFP — existing
Policy & approval orchestration
Role-based permissions and approval flows
Amount thresholds and address whitelisting
Time-window and velocity control rules
Static rule routing, single-request scope
Core question
Who is allowed to approve this?
RAFP+ — new AI layer
AI behavior risk engine
Dynamic scoring against 90-day user baseline
Platform-level anomaly detection across all accounts
Cross-account address convergence detection
LLM audit summaries generated in under 2 minutes
New question
Is this operation actually reasonable?
Live scenario

A real attack. Intercepted in real time.

Preview the full RAFP+ scenario.

RAFP+ Screening Engine · Quantum_Ex Debug Mode
POST /withdraw/apply
#001 16:51:24approved
user: 160000124 $840.66 USDT
#004 16:51:27hold
user: 160000128 $917.69 USDT
#016 16:51:33blocked
user: 160000005 $862.44 USDT
approved: 3 blocked: 17
Screening log
RISK_ENGINE profile_fetch user:160000124
history_days:214 new_addr_rate:8.1%
score:142/1000 action:APPROVED
————————————————————————————
PLATFORM concurrent_users:11 delta:+900%
PLATFORM new_addr_rate:100% baseline:8.1%
PLATFORM addr_convergence_confirmed
THmju... sources:[160000124, 160000005]
PLATFORM status:ELEVATED->CRITICAL
RISK_ENGINE freeze_accounts count:11
LLM incident_summary_ready
Analysis
User baseline loaded — 214-day history
daily_freq μ=0.6 new_addr_rate 8.1% script_prob 0.02
Platform layer — NORMAL → ELEVATED
3 concurrent accounts new_addr_rate 100% vs baseline 7.8%
CRITICAL — address convergence confirmed
THmju... ← [160000124, 160000005] interval: 7.4s
Run the full attack scenario