As companies rush to integrate AI into their workflows, security teams are facing a reality that is changing faster than traditional protection models can adapt. According to Francis de Souza, COO of Google Cloud, AI security can no longer be treated as an additional layer added later in development. It now sits at the center of enterprise architecture, governance, and operational strategy.

One of the biggest concerns highlighted by de Souza is the rapid growth of “shadow AI” inside organizations. Employees increasingly rely on consumer AI tools outside approved company systems, creating visibility and compliance gaps that most businesses struggle to monitor. The challenge extends beyond simple tool usage. AI systems introduce entirely new attack surfaces, including models, prompts, data pipelines, and autonomous agents that interact with internal systems.

The speed of modern attacks has also changed dramatically. De Souza noted that the average time between an initial breach and the next stage of an attack has fallen from several hours to just seconds. Traditional security workflows built around manual investigation and response are becoming increasingly difficult to sustain in environments where AI systems operate continuously and at machine speed.

Another emerging issue involves AI agents discovering forgotten or poorly secured internal systems. Older SharePoint servers, legacy storage environments, and outdated access permissions that previously stayed hidden inside organizations can suddenly become visible once AI agents begin navigating enterprise infrastructure autonomously. In practice, AI does not only accelerate productivity. It also accelerates exposure.

To address this, Google Cloud is advocating for what de Souza describes as “AI-native” defense systems, where defensive agents monitor, detect, and respond to threats with minimal human intervention. The role of leadership is also shifting. AI security is increasingly becoming a board-level concern rather than a problem handled exclusively by technical teams.

At the same time, the broader industry is still learning how to secure AI systems effectively. Lea Kissner, Chief Information Security Officer at LinkedIn, recently warned about what she called a coming “bug-pocalypse,” pointing to the growing shortage of professionals capable of managing AI-related vulnerabilities at scale.

What makes the conversation more complex is that even major platform providers are navigating these issues in real time. Recent investigations published by The Register revealed multiple cases where developers using Google Cloud unexpectedly received massive charges tied to unauthorized access to Gemini APIs. In several situations, API keys originally created for unrelated services like Google Maps had silently gained access to Gemini capabilities after permission scope changes.

The reports also highlighted concerns around billing controls and API revocation timing. Some developers discovered that deleted API keys could remain usable for several minutes due to delayed revocation propagation across infrastructure systems. Security researchers argued that newer authentication formats already demonstrate significantly faster revocation times, suggesting the issue may be more related to operational priorities than technical limitations.

The broader lesson for businesses is becoming increasingly clear: AI adoption introduces operational value, but also creates a new layer of infrastructure complexity that requires governance, visibility, and continuous oversight from the beginning. Security, architecture, workflows, permissions, and AI behavior are now deeply interconnected.

For companies building AI-enabled products or modernizing critical systems, the challenge is no longer simply adding AI capabilities. The real challenge is maintaining control, auditability, and resilience while systems become faster, more autonomous, and significantly more interconnected.

Source