Google Cloud’s CISO Chris Betz and senior director Ruchi Shah describe how the company has embedded specialized AI agents across its software development lifecycle (SDLC) to move from human-dependent checklists to autonomous, proactive security. The post argues that AI has made the traditional patching window obsolete, so defense must operate at machine speed.

Google Cloud’s architecture deploys modular, interconnected agents at every SDLC stage. In design, agents route product launches through a security review pipeline that cross-references designs against a continuous control catalog of over 200 requirements. High-risk indicators are automatically triaged for human intervention, and a dynamic product dossier replaces static threat models.

For code scanning, the team built Mantis, a multi-agent orchestration framework that avoids the low true-positive rates (under 7%) of naive AI scanning. Mantis constructs a hierarchical security summary tree, condensing files into directory and root-level summaries to reduce token overhead by over 85% while preserving structural context. Its workflow includes a Strategist agent that evaluates high-level code structure and threat models to create a prioritized investigation plan; Research agents that drill into raw source files; Deduplicator, reviewer, and critic agents that filter noise; and a Reproduction sandbox that automatically runs AI-generated proof-of-concept exploits in an isolated environment to verify exploitability before alerting developers. The core skills of Mantis have been open-sourced, with a more full-featured version running internally.

Dynamic fuzz testing is automated via an autonomous multi-agent engine that eliminates manual harness authoring. Context and Drafting agents synthesize product logic and unit tests to create initial fuzzing harnesses. Building and Testing agents execute code and feed compiler/linker errors to a Hallucination Cleaner agent that repairs broken dependencies. Quality Analyzer agents monitor runtime execution and adjust inputs to bypass blockers and penetrate deeper APIs.

Found vulnerabilities are routed into a unified patching pipeline: a Reproduce agent replicates the crash in sandbox, a Bug Context agent maps the failure path, a Patch agent generates a targeted fix, and an Evaluation agent runs a regression loop (recompile, test) to validate safety before submission to a human reviewer.

After deployment, an autonomous security posture management (ASPM) system converts the security standard catalog into programmable skills files, continuously checking production for configuration drift and triggering agentic remediation when violations occur.

A key cross-cutting component is a post-hoc self-reflection loop. After each workflow, a dedicated reflection agent analyzes execution logs, tool histories, and human feedback. Successful trajectories and design patterns are stored in a global knowledge store and injected into future agents’ context windows, creating a compounding improvement effect that the authors say has improved both vulnerability fix success rate and efficiency.

The article frames these internal practices as a path toward “immune” software that continuously discovers, validates, and patches its own weaknesses in real time. It also includes brief announcements (Confidential Computing updates, Sovereign Cloud, SecOps playbook, SIEM recognition, Wiz sensor, VPC Service Controls for agents, Gemini Spark bug hunting) and threat intelligence reports (UNC6508 campaign, ShinyHunters targeting Oracle PeopleSoft, zero-day in Cisco Catalyst SD-WAN Manager). It concludes with podcast recommendations.