Defense Layers
4 layers
Active
Input Scanning
Pattern + ML + Embedding + LLM Judge
3 layers
Active
Output Scanning
Credential, PII, harmful content
2 layers
Active
Deterministic Controls
Image stripping, URL allowlist
3 layers
Available
Deception Engine
Canary tokens, tarpit, fingerprint
Overview
Scans Today
0
across all scan types
Threats Blocked
0
block + sanitize actions
Input Scans
—
prompt injection detection
Output Scans
—
harmful content detection
Scan Breakdown
By Scan Type
Input Scans0 (0%)
Output Scans0 (0%)
RAG Scans0 (0%)
MCP Scans0 (0%)
Tool Validations0 (0%)
Detection Pipeline
1
Pattern Detection<3ms
700+ regex patterns, encoding detection, correlation signals
2
ML Keyword Classifier<1ms
TF-IDF weighted keyword scoring with sigmoid normalization
3
Embedding Similarity~5ms
Cosine similarity against 280+ known attack embeddings
4
LLM Judge~500ms
Claude Sonnet evaluates for social engineering and crescendo
Feedback
False Positives
0
reported by customers
Missed Attacks
0
reported by customers
Pending Review
0
awaiting triage
Applied
0
incorporated into detection
Automated Red Team
Daily
Run frequency
5 x 30
Rounds x attacks
287+
Known attack bank
4 layers
Input + Output + Embed + Determ.
Active
3 AM PT daily on Fly.io