Authors:
- Marina Kaganovich, Executive Trust Lead, Office of the CISO
- Anton Chuvakin, Security Advisor, Office of the CISO
First generative AI and now AI agents are bursting onto the scene, shaking up industries and opening doors to incredible innovation and efficiency. But let's be real: with all that power comes a huge responsibility. It's not just about building cool AI tools; it's about building them safely and securely. Here at Google, we're about helping businesses jump into this AI revolution head-first, while also making sure we're mitigating risks and creating applications you can truly trust. Think of it as balancing cutting-edge breakthroughs with a rock-solid foundation of secure and responsible innovation.
It's clear we're not alone in thinking this way. KPMG's latest report flagged data privacy as a top concern for 69% of business leaders diving into AI. So, yeah, everyone wants responsible AI, but getting there? That's the tricky part.
That's precisely why we put together Delivering Trusted and Secure AI, our latest paper outlining how Google approaches building enterprise-grade generative AI responsibly. We dive deep into how we handle AI data governance, privacy, security, and compliance, sharing our insights and best practices to help you navigate this complex landscape. Written by the experts in the Office of the CISO at Google Cloud, this paper tackles the questions we hear most often, so you can deploy AI with confidence and peace of mind.
Security
We believe security is absolutely foundational. That’s why our AI products are built atop a scalable technical infrastructure underpinned by a secure-by-design foundation and supported by robust logical, operational and physical controls to achieve defense in depth, at scale, and by default, designed to support you and accelerate secure, private, and compliant AI transformation.
In addition to building our AI products on a secure platform, we’ve also developed the Secure AI Framework (SAIF), a conceptual framework for securing AI systems across the four dimensions that make up an AI system: data, infrastructure, application, and model. SAIF is inspired by the security best practices – like reviewing, testing and controlling the supply chain – that we apply to software development, while incorporating our understanding of security megatrends and risks specific to AI systems.
Explore what changes and what stays the same when it comes to AI cybersecurity, best practices on securely deploying AI, and key considerations in applying SAIF throughout the AI development lifecycle.
Privacy
Privacy is a non-negotiable aspect of our approach to gen AI. One of the questions frequently posed to us is whether our foundation models are trained on customer data, and by extension, whether customer data may as a result be exposed to Google, Google’s other customers, or the public.
Our approach includes incorporating privacy design principles, designing architectures with privacy safeguards, and providing appropriate transparency and control over the use of data. When bringing new offerings to the market, we incorporate these principles throughout the product lifecycle and design architectures with comprehensive privacy safeguards such as data encryption.
Google Cloud’s enterprise customers have the ability to control their data. By default, Google Cloud does not use Customer Data to train its foundation models as articulated in the Google Cloud Platform Terms of Service and Cloud Data Processing Addendum. On our Vertex AI Platform, customer data prompted into a model and the output generated by Vertex AI from those prompts is “Customer Data,” and Google processes Customer Data only according to the customer’s instructions.
We offer a robust set of tools and services when it comes to protecting data and reducing the risk of customizing and training models with their own data that may include sensitive elements such as personal information (PI) or personally identifiable information (PII). To effectively segment, anonymize, and help protect data, consider implementing:
- Sensitive Data Protection, which provides sensitive data detection and transformation options such as masking or tokenization to add additional layers of data protection throughout a generative AI model’s lifecycle, from training to tuning to inference.
- VPC Service Controls, which allows for secure deployment within defined data perimeters where you can define perimeters to isolate resources, control and limit access, and reduce the risk of data exfiltration or leakage.
Compliance
Security and privacy in cloud computing are subject to legal, regulatory compliance, and risk management requirements. To provide customers with transparency into our controls, we make compliance documentation, certifications, control attestations, and independent audit reports readily available to satisfy regional and industry-specific requirements and compliance validation efforts.
We are committed to GDPR compliance and have compiled comprehensive DPIA Resource Center documentation to support you in performing data protection impact assessments. We also enable certain AI/ML services to be configured to meet data residency requirements as noted in our Service Terms.
We understand that AI comes with complexities and risks, and to ensure our readiness for the future landscape of AI compliance we proactively benchmark ourselves against emerging AI governance frameworks. To put our commitments into practice, we invited Coalfire to examine our current processes, measure alignment and maturity toward the objectives defined in the National Institute of Standards and Technology (NIST) Artificial Intelligence Risk Management Framework (AI RMF) and International Organization for Standardization (ISO) ISO/IEC 42001 standard.
Coalfire’s assessment provided valuable insights, allowing us to enhance our security posture as we continuously work to uphold the highest standards of data protection and privacy. We believe that an independent and external perspective offers critical objectivity, and we are proud to be among the first organizations to perform a third-party AI readiness assessment.
Looking Ahead
We aim to be at the forefront of advancing AI through our deep research to develop more capable and useful AI, ensuring effective risk management strategies evolve along with it. We recognize that this requires an industry-wide effort, which is why we co-founded the Coalition for Secure AI (CoSAI), to support collective investment in AI security through collaboration amongst industry and academia.
AI continues to evolve, AI agents are taking center stage and securing them becomes a priority. The path to trusted and secure AI isn't just a Google endeavor—it's a shared journey. This increasing capability and autonomy promises significant value, potentially reshaping how businesses operate and individuals interact with technology. Ensuring we innovate responsibly is critical.
In recent months, we’ve proudly announced the Agent2Agent Protocol and are thrilled to share our approach to securing AI agents. For further reference, also check out CSA’s AI Controls Matrix and OWASP’s Gen AI Security Project for practical strategies and implementation resources such as the Top 10 Risks and Mitigations for 2025.
