AI Security Detectors

Oculum is purpose-built for AI-native security. While traditional security scanners focus on OWASP-style vulnerabilities, Oculum specializes in detecting security issues unique to LLM-powered applications.

Why AI Security Matters

AI applications introduce new attack surfaces that traditional tools miss:

Prompt Injection — Malicious input that manipulates LLM behavior
Unsafe Execution — AI-generated code executed without validation
Data Exfiltration — Sensitive data leaked through AI responses
Agent Takeover — Autonomous agents performing unintended actions
Supply Chain — Compromised models or packages

Oculum's detectors are designed specifically for these AI-era threats.

Detection Categories

AI-Specific Detectors

Category	What It Detects	Severity
Prompt Injection	Unvalidated user input in LLM prompts	High-Critical
Unsafe Execution	AI-generated code executed without checks	Critical
Package Hallucination	Non-existent packages suggested by AI	High
RAG Security	Data exfiltration, poisoning, PII leaks in RAG	Medium-Critical
MCP Security	Tool poisoning, credential exposure, prompt injection	High-Critical
Agent Safety	Excessive agency, overpermissive tools	Medium-High
Model Supply Chain	Unsafe model loading, unverified sources	High-Critical

Traditional Detectors

Category	What It Detects	Coverage
Secrets	API keys, tokens, credentials	Comprehensive
Config Audit	Debug flags, insecure settings	Standard
Weak Crypto	MD5, SHA1, DES usage	Standard

Three-Layer Detection

Oculum uses a layered approach for accurate detection:

Layer 1: Pattern Matching

Fast, deterministic detection:

Regex patterns for known secrets
Entropy analysis for potential secrets
URL detection for sensitive endpoints

Speed: ~1000 files/second

Layer 2: Structural Analysis

Context-aware heuristics:

Auth pattern recognition
Dangerous function detection
AI code fingerprinting
Data flow tracking

Speed: ~500 files/second

Layer 3: AI Semantic Analysis

Deep understanding using Claude:

Full taint analysis
Business logic validation
Cross-file flow tracking
Framework-aware analysis

Speed: ~50 files/second

Severity Levels

Level	Meaning	Action
Critical	Immediate exploitation risk	Fix before deploy
High	Significant security risk	Fix soon
Medium	Moderate risk, needs review	Review and decide
Low	Minor issues	Consider fixing
Info	Informational	Awareness only

Confidence Levels

Each finding includes a confidence level:

High — Strong evidence, low false positive rate
Medium — Good evidence, may need context review
Low — Potential issue, requires manual verification

AI-validated findings (--depth verified) have higher confidence levels.

Filtering by Category

Fail on Specific Categories

# Only fail on AI-related issues
oculum scan --fail-on-categories "ai-*"

# Only fail on secrets
oculum scan --fail-on-categories "secrets-*"

# Multiple categories
oculum scan --fail-on-categories "ai-*,secrets-*"

Available Category Prefixes

ai-* — All AI security detectors
secrets-* — Secret detection
owasp-* — Traditional web vulnerabilities
config-* — Configuration issues

Understanding Findings

Each finding includes:

Category — The detector that found it
Severity — Impact level
Confidence — Detection certainty
Location — File, line, column
Description — What was found
Remediation — How to fix it
Validation Notes — AI reasoning (for verified scans)

Example Finding

{
  "category": "ai_prompt_injection",
  "severity": "high",
  "confidence": "high",
  "message": "User input passed directly to LLM prompt",
  "file": "src/api/chat.ts",
  "line": 45,
  "remediation": "Validate and sanitize user input before including in prompts",
  "validationNotes": "User input from request body is concatenated directly into the system prompt without sanitization."
}

Quick Reference

| If you're building... | Focus on these detectors | |----------------------|-------------------------| | Chatbot / AI Assistant | Prompt Injection, Agent Safety | | RAG Application | RAG Security, Data Exfiltration | | Code Generation Tool | Unsafe Execution, Package Hallucination | | MCP Server/Client | MCP Security, Tool Poisoning | | Fine-tuning Pipeline | Model Supply Chain |

Next Steps

Explore individual detector categories:

Prompt Injection — Protect against prompt manipulation
Unsafe Execution — Secure AI-generated code
RAG Security — Secure retrieval-augmented generation
MCP Security — Secure Model Context Protocol
Agent Safety — Control autonomous agents