AI Security: How Companies and Google Adapt in Real Time

AI adoption demands a fundamental shift in cybersecurity strategies. As Google Cloud COO Francis de Souza noted, security cannot be an afterthought or rely on individual employee initiatives. A major risk is "shadow AI"—employees using unauthorized tools outside corporate standards. De Souza stressed that AI development strategies must integrate data protection, access management, and audit capabilities from the outset.

Modern threats evolve faster than traditional defenses. The window between vulnerability discovery and attack initiation has shrunk from eight hours to just 22 seconds. The attack surface has expanded to include not only networks but also AI models, training data pipelines, agents, and even prompts. De Souza highlighted a frequently overlooked risk: AI agents can uncover forgotten data repositories, such as outdated SharePoint servers with long-expired access controls. These legacy systems create vulnerabilities, as attackers may exploit them to steal sensitive information.

De Souza advocates for fully automated security systems. Instead of relying on manual processes or partial automation, companies should deploy AI-driven solutions where agents autonomously detect and neutralize threats. Yet even with advanced technology, the challenge remains a matter of corporate governance. Security, he emphasized, is not solely an IT department responsibility—it requires engagement from boards of directors and top management.

Real-world incidents underscore the persistent security challenges platforms face. Recent cases involving Google Cloud developers receiving unauthorized bills exceeding tens of thousands of dollars due to unchecked API access to Gemini revealed flaws in access management. The platform automatically increased spending limits without explicit user consent, leading to financial losses. Research by Aikido also found that even after revoked API keys were deleted, attackers could still exploit them for up to 23 minutes due to delayed access revocation propagation in Google’s infrastructure. While newer Google credential formats address this issue more efficiently, the company has yet to implement widespread fixes.

The situation is further complicated by a shortage of AI security professionals. LinkedIn’s Chief Information Security Officer, Lee Kissner, noted that the industry lacks long-term solutions for AI vulnerability management. In this environment, businesses must not only follow platform recommendations but critically assess their own preparedness for emerging threats.