At a recent event in Los Angeles, Google Cloud COO Francis deSouza shared critical insights into where enterprises stand on AI security. According to deSouza, the tech industry is in a challenging transition period—one that even Google itself is actively navigating in real time.
DeSouza’s core message is clear: security cannot be an afterthought. "As companies embark on this AI journey, they need to take a platform approach," he urged. He specifically warned against "shadow AI"—employees using consumer-grade AI tools without organizational oversight. He emphasized that an AI strategy is completely inseparable from a robust data and security strategy; they must go hand in hand.
Importantly, deSouza advocated for a consistent, multicloud security posture rather than pitching Google Cloud in isolation. He noted that even companies aiming for a single-cloud environment inevitably operate across multiple clouds due to third-party SaaS integrations and business partners. Therefore, consistent security across different clouds and models is mandatory.
Furthermore, the threat landscape has accelerated beyond traditional defense speeds. DeSouza revealed a startling statistic: the average time between an initial breach and the next stage of an attack has plummeted from eight hours to just 22 seconds. The modern attack surface has expanded well beyond network perimeters to encompass AI models, training data pipelines, prompts, and active agents.
One critical but overlooked threat is the risk of "roaming agents." DeSouza explained that autonomous agents navigating internal networks will easily discover forgotten, unpatched data repositories—like legacy SharePoint servers—that survived previously only due to obscurity. These agents will rapidly expose hidden sensitive assets.
The solution, in his view, is meeting machine speed with machine speed. We are seeing the emergence of an AI-native, fully agentic defense where enterprises deploy autonomous security agents to run real-time, automated defenses.
[AgentUpdate Depth Analysis] The rapid onboarding of AI agents introduces a systemic paradigm shift in enterprise security. Legacy Role-Based Access Control (RBAC) was designed for human limitations, not for autonomous agents capable of indexing and synthesizing massive, forgotten internal data silos in seconds. By making "security through obscurity" obsolete, roaming agents force enterprises to transition to real-time, context-aware auditing of tool calls, prompt injections, and RAG pipelines. Moving forward, the industry will see the rise of dedicated "Agentic Security" frameworks. Defending enterprise boundaries will require deploying defensive AI agents to counter offensive machine-speed threats. This shift will redefine standard interface protocols like MCP (Model Context Protocol), positioning automated guardrails at the core of the future enterprise AI stack.
