Google Cloud CISO Chris Betz outlines four lessons learned while building AI Threat Defense, a framework for using AI to defend against AI-powered attacks. The post argues that AI has rewritten cybersecurity rules: attackers now exploit not just source code but configurations, binaries, and firmware at machine speed, but defenders gain an equal advantage when armed with full business context and AI tools. The framework—Prepare, Scan and Prioritize, Remediate, Monitor—is based on internal projects like Project Naptime, Big Sleep, and CodeMender.

1. Prepare: Reduce the attack surface to focus efforts. Invest in an operational scaling framework that tracks software supply chains, dependency graphs, and token budgets. Align engineering and security work by tagging scanned components, allocating hardware and token budgets, and managing change volume with small, safe rollouts.

2. Scan and Prioritize: Continuously scan across products (Search, Ads, Android, Chrome, Google Cloud). The best results come from combining a product expert, a good harness, and an AI model. If only two are possible, choose expertise and harness—a weaker model with a strong harness outperforms a strong model alone. Use multiple models, iterate on data, engage key suppliers, and flip the usual prioritization: tackle foundational code with the largest blast radius first, not the smallest.

3. Remediate: Adopt a risk-based patch rollout that prioritizes externally reachable code with broad impact (e.g., BoringSSL, gVisor). Provide models with full context for faster fixes. Build a central tracking system with service-level objectives (SLOs) and deploy constant autonomous patching with human review and rollback capabilities. The ultimate goal is a resilient system that patches itself, not just a patching system. Extend this to open source via a three-R strategy: Refresh (fix foundational code), Remove (replace peripheral dependencies), and Rewrite (transition legacy logic to memory-safe languages using AI).

4. Monitor: Create a perpetual feedback loop that tracks system strain and vulnerability hotspots. Build a comprehensive asset inventory with product-level SLOs for vulnerability management. Use AI agents to automate response playbooks, verify patch efficacy, and feed key libraries into Gemini for improved pattern recognition. Red teams continuously stress-test core infrastructure. The outcome is a living, measured program that hardens over time.

The post concludes by positioning AI Threat Defense as a pathway any team can use to build a learning, adaptive defense against future threats.