Security, IP & Team Adoption

Security: what goes into prompts

Every prompt you send may be logged, stored, or used for training (depending on provider and plan). Treat prompt content as data classification:

Classification	Can send to AI?	Example
Public	Yes	Open-source code, public docs, Stack Overflow
Internal	Yes, on enterprise plan	Internal APIs, architecture docs, non-secret code
Confidential	Only with approved tools + DPA	Customer data, unreleased features, business metrics
Restricted	Never	Credentials, PII, encryption keys, security vulns

Enterprise/API plans (OpenAI, Anthropic) typically offer: no training on your data, configurable retention, SOC 2 compliance. Consumer products (ChatGPT free, Claude free) may use data for training. Know which you're using.

Secret leakage prevention

The #1 security incident vector for AI-assisted development:

Vector	Prevention
Pasting .env files into prompts	.cursorignore + rules: "never paste secrets"
AI hardcodes API keys in generated code	Pre-commit hook scanning for secrets (gitleaks, trufflehog)
Agent reads credential files via MCP	Scope MCP servers; exclude secrets from indexing
Prompt logging in CI	Sanitize logs; don't log full prompts with secrets
Copy-paste from prod logs	Scrub PII/secrets before pasting; rules against raw log paste

Add to your rules: "Never include credentials, API keys, tokens, or PII in prompts. If you need to reference a config value, use placeholder names."

IP ownership of AI output

Legal landscape (mid-2026, varies by jurisdiction — consult your legal team):

Question	Typical enterprise position
Who owns AI-generated code?	The company (work for hire, same as human output)
Can AI output infringe patents/copyright?	Possible — review for copied code (GitHub Copilot litigation precedent)
Must AI usage be disclosed?	Depends on company policy and jurisdiction
Can you open-source AI-generated code?	Only if human review confirms no license conflicts

Practical policy: AI output goes through the same IP review as human code. Run license scanners on AI-generated PRs. Don't assume "AI wrote it" means "we own it cleanly."

Enterprise AI governance framework

For teams scaling AI adoption:

Policy layer: What tools are approved? What data classifications are allowed? ↓ Configuration: Shared rules, skills, model routing standards ↓ Workflow: Review requirements, eval gates, automation guardrails ↓ Monitoring: Usage metrics, cost tracking, incident response ↓ Review cycle: Quarterly policy review, model upgrade evals

Governance element	Owner	Frequency
Approved tools list	Eng leadership + Security	Quarterly
Data classification policy	Security/Legal	Annually
Shared rules/skills	Team leads	Ongoing (PR review)
Cost budget	Eng manager	Monthly review
Eval suites for production agents	Feature team	Per release
Incident response (AI-specific)	On-call + Security	As needed

Team adoption patterns that work

Top-down mandates fail. What works: show real wins with metrics, weekly office hours, shared config in git, pairing sessions (experienced user + skeptic), and approved enterprise tools with clear data policies.

Anti-pattern	Why it fails
Mandate usage quotas	Gaming metrics; resentment
Ban AI entirely	Shadow usage with consumer tools (worse security)
No policy	Inconsistent practices; security incidents
Only track cost	Can't justify investment

Incident response for AI-specific issues

Incident	Response
Secret pasted into prompt	Rotate credential; review provider data retention policy
AI generates vulnerable code merged to prod	Standard security incident + add eval case
AI automation runs uncontrolled	Kill switch for automation; review guardrails
Wrong data sent to AI provider	Data breach assessment per company policy

Document these in your team's incident playbook alongside traditional categories.

Interview framing

"I establish data classification for prompts (public/internal/confidential/restricted), use enterprise plans with no-training guarantees, scan AI output for secrets and license issues, and treat AI config as governed code in git. Adoption works through demonstration and shared config, not mandates."

Senior signal: Reference your company's or a realistic data classification policy. Explain why banning AI is worse than governing it. Describe a specific governance element you'd implement first (usually: data classification + enterprise plan).