Google Cloud COO flags AI security

Google Cloud COO Francis de Souza said on May 24 that even Google is still working through how to secure increasingly autonomous AI systems as the company rolls out new agentic products. In an interview published by TechCrunch, de Souza said companies are “navigating AI security in real time” as they move from chat-style tools toward software that can take actions on a user’s behalf. TechCrunch said the interview took place backstage at an event in Los Angeles. The remarks came days after Google used its I/O 2026 conference to promote a broader push into what it called AI that can act, not just answer. ### What exactly did Francis de Souza say? TechCrunch quoted de Souza as saying, “We’re in the transition period — all of us,” describing a period in which security practices for AI are still being built out. The publication said he added that companies are effectively learning in public as they deploy more capable systems. TechCrunch framed the point as notable because it came from a senior Google executive while Google itself is introducing more agentic tools. (techcrunch.com) Francis de Souza has been speaking publicly about AI security in other Google venues as well. In an April 22 Google Cloud blog post tied to Cloud Next 2026, de Souza wrote about “cybersecurity in the era of the agentic enterprise” and described a shift toward securing agents, AI-generated code and what he called the “agentic web.” That post did not use the same wording as the TechCrunch interview, but it showed Google presenting AI security as an active operational issue rather than a finished framework. (techcrunch.com) ### Which Google products made the security question more urgent? Google said at I/O 2026 that Gemini 3.5 Flash combines “frontier intelligence with action,” while Gemini Omni is designed for multimodal generation and editing. Google also said it had moved “beyond AI tools that just help us write, to agents that help us act,” language that underscored the company’s push toward systems that can carry out tasks. (cloud.google.com) Google Cloud used similar language in its I/O 2026 enterprise announcements. In a company blog post, Google said it was “doubling down” on support for the “Agentic Enterprise” through Gemini Enterprise, Agent Platform and Workspace products. That matters because the more an AI system can act across business software, the more questions arise around permissions, logging, review and misuse. That connection was made explicitly in Google’s own security materials for Cloud Next 2026. (blog.google) ### What security controls has Google said it is building? Google Cloud said in April that it was adding products and features meant to secure AI deployments, including Model Armor integration with Agent Gateway, Agent Identities, AI-BOM tools for AI-generated code and expanded Wiz coverage for AI and cloud apps. The company also said new agents in Google Security Operations could help with threat hunting, detection engineering and third-party context. (cloud.google.com) Those announcements show Google pairing its agent push with security tooling, but they also show the controls are still being assembled as products roll out. TechCrunch’s interview with de Souza put that more plainly by acknowledging that guardrails, auditing and broader operating practices are still evolving alongside the technology. (cloud.google.com) ### Why does this matter now, rather than later? Google made the timing clear at I/O 2026 by presenting agentic features as current products rather than distant research. The company’s public materials described AI systems that can act across enterprise workflows and consumer experiences, while de Souza’s comments indicated the security model for those systems is still being worked through in practice. (techcrunch.com) TechCrunch published the de Souza interview on May 24, 2026. Google’s related roadmap is already public in its I/O 2026 and Cloud Next 2026 materials, where Gemini Enterprise, Agent Platform, Agent Gateway and security products were named as the next operating layer around those systems. (techcrunch.com) (blog.google)