How customer collaboration is shaping the future of GenAI security with Model Armor

At Google Cloud, we believe that the best products are built in partnership with our customers. Their feedback and real-world experiences are invaluable in helping refine our services and deliver solutions that truly meet our customers’ needs. In January 2026, our Google Cloud Developer Advocacy team participated in a high-velocity technical sprint with a major Google Cloud customer and a leader in the telecommunications industry.

This collaborative engagement provided us with deep insights, leading to significant enhancements in Model Armor information experience, our service for Runtime security for generative and agentic AI.

Accelerating GenAI adoption through "radical empathy"

The objective of this engagement was to support the productionization of a next-generation GenAI customer support platform built using Google Cloud's Agent Development Kit (ADK) and Agent Platform. By sitting directly with the customer's developers and security specialists, we gained a unique opportunity to observe how developers interact with Gemini Enterprise Agent Platform in a live, complex environment.

This experience provided something traditional documentation cycles cannot replicate: radical empathy. By logging friction points, as developers worked, we translated functional blockers into technical insights in real-time, identifying exactly where developers were hindered by ambiguous configuration guidance or a lack of granular detail.

Key discoveries from the front lines

By observing the development workflow firsthand, we identified four critical friction points:

• Search-first workflows: Developers rarely navigate through documentation hierarchies; instead, they rely on search to jump straight to specific code examples. A lack of comprehensive, copy-pasteable snippets for common use cases—like PII redaction—was a primary point of friction.
• Balancing confidence levels: Finding the right balance between comprehensive threat detection and minimizing disruptive false positives proved challenging. For instance, using aggressive settings like "low and above" often caused a high volume of false positives that interrupted legitimate customer support flows.
• The need for granular guidance: While the core concepts of Model Armor were understood, developers needed more detail on how different enforcement methods function in practice to balance security with usability.
• Integration roadblocks (the 403 error): When integrating Model Armor with other services like Apigee, developers frequently encountered 403 PERMISSION_DENIED errors. This indicated a gap in our documentation regarding necessary cross-service IAM roles and permissions.

Turning insights into action

The insights gained from this partnership were immediately channeled into a comprehensive overhaul of Model Armor’s documentation and guidance:

• Tested, copy-pasteable code samples: We have added numerous tested, ready-to-use code samples throughout the documentation to support search-first workflows.
• The confidence level matrix: We introduced a new technical reference to help users understand the trade-offs between different filter levels. We now explicitly recommend "High" or "Medium" thresholds for general content to minimize false positives, reserving "Low and above" for high-security threats like prompt injection and jailbreak detection.
• Explicit integration guides: We updated our integration guides, with a focus on Apigee, Gemini Enterprise Agent Platform, and GKE. These now clearly outline the specific IAM roles required (such as roles/modelarmor.user) to ensure smooth, error-free deployments.
• Deeper technical documentation: We have enhanced the documentation to provide in-depth explanations of enforcement methods and their real-world applications.

The power of partnership

Getting "in the room" with our customers allowed us to bridge the gap between technical accuracy and operational utility. This journey of co-innovation ensures that Model Armor serves as a genuine catalyst for your success. We encourage you to explore the updated documentation and share your feedback as we continue to build the most secure platform for your GenAI workloads.

Get started:

• Explore the updated Model Armor documentation