AI Security Static Engine
Deterministic static code analysis for AI/ML attack surface detection. AST-based scanning with 82+ core rules plus 5 specialized profiles for prompt injection, agent safety, data leakage, and tool abuse.
What It Does
Static Code Analysis
Analyzes source code using AST (Abstract Syntax Tree) parsing to detect AI-specific vulnerabilities without executing code.
Prompt Injection Detection
Identifies user input reaching LLM prompts without validation (R1). Detects 15+ patterns including direct concatenation and template injection.
Data Leakage Prevention
Detects sensitive data exposure through LLM outputs, logging, and error messages (R7.1-R7.6). Tracks data flow from source to sink.
Compliance Evidence
Maps findings to SOC2, GDPR, HIPAA, ISO27001, and OWASP LLM Top 10. Generates cryptographically signed artifacts for auditors.
What It Doesn't Do
Important limitations to understand:
No Runtime Testing
Cannot detect vulnerabilities that only appear at runtime. Use Runtime Engine for that.
No AI/ML Detection
Uses deterministic pattern matching, not machine learning. 100% reproducible results.
No Code Execution
Analyzes code structure only. Never runs your code or makes external requests.
No Certification
Provides evidence for auditors. Does not certify compliance or security.
Quick Start (5 minutes)
Connect Your Repository
Navigate to Dashboard → AI Security → New Scan and enter your GitHub repository URL.
https://github.com/your-org/your-ai-appRun Scan
Click "Start Scan". The engine will clone your repo (ephemeral), analyze it, and delete the clone.
Privacy: Repository is cloned to /tmp/scan_[timestamp] and deleted immediately after scan. No code is stored.
Review Findings
View findings categorized by severity (Critical, High, Medium, Low) with fix recommendations.
Try Interactive DemoCore Rule Families (R1-R14)
Prompt Injection
User input reaches system/developer prompt without validation
Tool Abuse
Model output triggers privileged action (PythonREPL, Shell, eval/exec)
Tool Argument Validation
Tool arguments lack validation or sanitization
RAG Poisoning
Indirect prompt injection via unvalidated retrieval documents
Auth / Access Control
Missing authentication on AI entrypoints
Multi-Tenant Isolation
Missing tenant scoping on AI-accessible data
Agent Safety Suite
Agent loops, memory, function calling, parallel execution, streaming, multi-modal input
Secrets Exposure
Secrets flow to AI context, logs, or outbound channels
SSRF / Egress Control
Outbound requests lack strict egress controls
AI Control Stability
Non-deterministic AI on privileged execution paths
Authorization / RBAC
Missing AI authorization boundary enforcement
Output Validation
AI output lacks validation before downstream use
Resource Protection
Missing rate limiting on AI endpoints
Context Window Security
Context window overflow risk
Model Confusion
Model confusion or parameter injection attacks
Specialized Scan Profiles
Profiles are additive — selecting multiple profiles merges their rule packs with deduplication. The DEFAULT profile is always included.
Default
R1-R14Core AI security rules covering prompt injection, tool abuse, secrets, SSRF, agent safety, and more
Voice AI
VAI-Telephony webhook security, STT input boundaries, session integrity, PII/recording compliance
Agentic Workflow
AGW-RAgent loop safety, tool selection governance, action governance, state integrity, idempotency
Embedded SaaS
ESA-RRBAC enforcement, context minimization, stored injection, feature gating, audit logging
Deep RAG
RAG-RRetrieval scope, embedding access control, document provenance, embedding drift detection
AI-Assisted Dev
AIC-RSupply chain integrity, IAM risk, license/IP contamination, testing enforcement, AI governance