As companies rush to adopt AI, Google Cloud’s COO Francis de Souza warns that security must be built in from the start, not added later. He emphasizes that AI expands the attack surface with models, data pipelines, and autonomous agents, making traditional defenses too slow. The solution, he argues, is AI-driven, fully agentic defense systems overseen by leadership, not just security teams.

However, Google’s own platform has recently faced criticism. Developers using Google Cloud have seen five-figure bills from unauthorized Gemini API calls, due to expanded API key access and automatic billing tier increases without consent. Google refunded the charges but stands by its policy, prioritizing service continuity over budget controls.

More critically, research from security firm Aikido shows that even when developers revoke compromised API keys, Google’s revocation takes up to 23 minutes to propagate. During that window, attackers can still authenticate over 90% of requests, exfiltrating data. Newer Google credential formats revoke in seconds, suggesting the delay is a matter of prioritization, not technical limits.

The gap between Google’s security guidance and its own platform behavior highlights a broader industry challenge: even the leaders in AI are still adapting. As AI agents uncover forgotten data stores and attack speeds accelerate, organizations must demand consistent, cross-cloud security — and hold providers accountable when they fall short.