This document describes the security measures implemented in Cortex and how to maintain them.

All tool inputs are validated before processing:

Protection Against:

• Command injection
• Path traversal
• Buffer overflow
• Invalid type coercion

Prevents runaway API costs with tiered limits:

Features:

• Sliding window algorithm
• 1.5x burst allowance
• 60-second cooldown on limit breach
• Environment variable override: CORTEX_RATE_LIMIT=false

All operations are logged to JSONL files:

Log Location: ~/.claude/memory/logs/cortex-YYYY-MM-DD.jsonl

Logged Events:

• tool_call_start / tool_call_end / tool_call_error
• rate_limit_hit
• validation_failure
• resource_access
• prompt_access
• session_start / session_end

Features:

• Automatic log rotation (10MB max)
• Keeps last 5 rotated files
• Sensitive field redaction
• Correlation IDs for request tracing

Environment Variables:

• CORTEX_AUDIT=false - Disable logging
• CORTEX_LOG_LEVEL=DEBUG|INFO|WARN|ERROR - Set log level
• CORTEX_AUDIT_CONSOLE=true - Echo to stderr

Optional encryption for sensitive memory data:

Algorithm: AES-256-GCM (authenticated encryption) Key Derivation: PBKDF2 with SHA-512, 100,000 iterations

Setup:

# Generate a secret
node -e "console.log(require('./core/encryption.cjs').generateSecret())"

# Set the secret
export CORTEX_ENCRYPTION_SECRET="your-base64-secret"

Usage:

const { encrypt, decrypt } = require('./core/encryption.cjs');

const encrypted = encrypt('sensitive data');
const decrypted = decrypt(encrypted);

1. npm audit - Check for known vulnerabilities

   cd ~/.claude/memory
   npm audit

2. snyk - Deep dependency scanning

   npm install -g snyk
   snyk test

3. eslint-plugin-security - Static code analysis

   npm install --save-dev eslint-plugin-security
   # Add to .eslintrc: plugins: ['security']

4. retire.js - Detect outdated libraries

   npm install -g retire
   retire --path .

Run periodically (recommended: weekly):

• Dependencies: Run npm audit and update vulnerable packages
• API Key: Verify ANTHROPIC_API_KEY is not hardcoded
• Log Files: Check logs don't contain sensitive data
• Permissions: Verify file permissions on ~/.claude/memory/
• Rate Limits: Review rate limit stats for anomalies
• Encryption: Verify encryption secret is not committed

Trusted:

• Claude Code client
• Local file system
• Environment variables

Untrusted:

• Tool input arguments (always validated)
• Resource URIs (resolved and checked)
• Prompt arguments (sanitized)

If you suspect a security issue:

1. Disable the server: Remove from ~/.claude.json
2. Review logs: Check ~/.claude/memory/logs/
3. Rotate API key: Generate new key in Anthropic console
4. Rotate encryption secret: Generate new secret, re-encrypt data
5. Report: File issue at repository (do not include sensitive data)

• v1.0.0: Initial security implementation
  • Input validation for all 6 tools
  • Rate limiting with sliding window
  • JSONL audit logging with rotation
  • Optional AES-256-GCM encryption