Privacy Deep Dive

Detailed technical analysis of privacy implications.

What Gets Sent

When you use AI coding tools, they typically send:

• Current file content
• Recent file history
• Codebase embeddings (if indexed)
• File paths and structure
• Your prompts
• Git diff (sometimes)
• Terminal output (sometimes)

Jurisdiction Risk

Jurisdiction	Legal Framework	Risk Level
US	CLOUD Act, FISA 702	Government can compel disclosure
EU	GDPR	Stronger user protections
China	PIPL, Cybersecurity Law	Data localization requirements

US CLOUD Act

Allows US government to compel disclosure of data stored by US companies anywhere globally, potentially without user notification.

Implications

• Data stored by US company = subject to US law
• Applies even if servers are in EU
• Enterprise agreements may provide some protection

Verification Methods

What You CAN Verify

• Network destinations (where data goes)
• Payload contents (what’s sent)
• Local file access (what’s read)

What You CANNOT Verify

• Server-side retention (must trust vendor)
• Training exclusion (must trust vendor)
• Government access (no visibility)

Tools for Verification

# Monitor outbound connections
sudo lsof -i -n | grep cursor

# Use mitmproxy for traffic inspection
mitmproxy --mode regular

# Check request headers for privacy mode
# Look for: x-ghost-mode: true

API Key Security

Practice	Recommendation
Storage	System keychain, not env files
Rotation	Quarterly minimum
Scope	Separate keys for AI vs production
Git	Pre-commit hooks with gitleaks

Secure Key Storage

# macOS Keychain
security add-generic-password -a "$USER" -s "openai-api-key" -w "sk-..."
export OPENAI_API_KEY=$(security find-generic-password -s "openai-api-key" -w)

# Linux (GNOME Keyring)
secret-tool store --label="OpenAI" service openai username apikey
export OPENAI_API_KEY=$(secret-tool lookup service openai username apikey)

Threat Models

Low Risk (Personal Projects)

• Use any tool with privacy mode
• Basic exclusions for secrets

Medium Risk (Client Work)

• Use tools with zero retention
• Comprehensive exclusions
• Review before each session

High Risk (Regulated Industry)

• Self-hosted or local only
• Air-gapped where possible
• Formal security review

Security Incidents to Know

Incident	Date	Lesson
Anthropic training default ON	Aug 2024	Check defaults
Package hallucination attacks	Ongoing	Verify dependencies
Prompt injection demos	Ongoing	Limit agent permissions

Minimizing Exposure

1. Enable privacy mode immediately
2. Create comprehensive exclusions
3. Use BYOK where possible
4. Prefer local models for sensitive work
5. Review diffs before committing
6. Rotate keys regularly