How to secure AI agents at machine speed: Google Cloud's agentic security stack

As autonomous agents move into production pipelines, security teams face a new reality: machine-speed commits, unvetted dependencies and over-privileged credentials can turn AI agents into direct backdoors. Human review no longer scales to the velocity of agentic development, and flat-trust architectures expose enterprises to lateral movement and data exfiltration. At the Google Cloud Security Summit, Google presented a unified framework to build securely, secure workloads and defend at machine speed. Softprom is the official partner of Google.

What was announced

Google Cloud introduced updates across three layers of agentic security: secure development, runtime governance and autonomous defense. CodeMender, the autonomous code-security agent, is now in preview with secure sandboxed remediation and flexible scanning modes. It works with multiple models including Gemini 4 Argon, which recently helped Wiz identify core PHP vulnerabilities affecting 70% of websites. Since July, hundreds of customers have adopted CodeMender, with early evaluations showing vulnerability discovery time reduced by over 95%.

Auth Manager is now generally available, centralizing OAuth 2.0/PKCE credential management in a Google-managed vault. Agent Identity JWT support (preview) issues OIDC tokens so agents can authenticate across AWS, Azure and on-premises backends without long-lived API keys. VPC Service Controls now support IAM roles (GA) and Workload Identity Federation for GKE (preview). Model Armor adds a custom topic detector (preview) for healthcare, insurance and retail use cases. Cloud NGFW gains an advanced malware sandbox powered by Palo Alto Networks WildFire.

Why this matters

IDC research cited in the announcement finds that nearly 40% of enterprises add vulnerabilities faster than they can remediate them. For CIOs, CISOs and procurement leaders, the shift to agentic workflows amplifies this gap: autonomous coding agents commit code faster than human reviewers can audit, while runtime agents expand the attack surface across multi-cloud environments. Google Cloud's response pairs exploitability validation with deployable fixes, enforces identity-based perimeters and extends native controls to multi-cloud deployments through the deepened Wiz partnership.

As adversaries attack at machine speed, Google's Fairwind Program helps us keep pace with our application security. Our early evaluation of CodeMender shows the potential to reduce vulnerability discovery time by over 95%, while keeping expert validation and remediation under human oversight

Technical details

• CodeMender: autonomous remediation inside OS-level sandbox or isolated VM, human-in-the-loop for write and execute, scan by folder, repo or modified files.
• Auth Manager (GA): OAuth 2.0/PKCE automation, Google-managed vault, IAM-governed refresh tokens.
• Agent Identity JWT (preview): OIDC tokens for AWS, Azure, on-prem and MCP servers.
• VPC Service Controls: IAM-role ingress/egress (GA), Workload Identity Federation for GKE (preview).
• Model Armor custom topic detector (preview): prompt-based on-topic/off-topic rules.
• Cloud NGFW: advanced malware sandbox with Palo Alto WildFire (GA), project-level resources (GA), Cloud Armor managed rules (preview) backed by Thales Imperva threat intelligence.
• Wiz integrations: Wiz AI SAST (preview), Wiz Defend and Google Security Operations shared data model, Google Threat Intelligence in Wiz.

FAQ

How does CodeMender reduce vulnerability discovery time?

CodeMender uses agentic reasoning with models including Gemini 4 Argon to identify unknown vulnerabilities, generate tested patches and validate fixes in isolated sandboxes. Thales reported over 95% reduction in discovery time during early evaluation.

Can autonomous agents authenticate across multi-cloud environments?

Yes. Agent Identity JWT support, in preview, issues standard OIDC tokens that allow agents to authenticate securely to AWS, Azure, on-premises backends and MCP servers without static long-lived API keys.

How does Google Cloud prevent agents from going off-topic?

Model Armor's custom topic detector, available in preview, lets teams define prompt-based rules for on-topic and off-topic interactions, tailored for healthcare, insurance and retail compliance requirements.

Softprom and Google

Softprom is the official partner of Google. Our team helps enterprises design, deploy and operate Google Cloud security architectures, including CodeMender, Auth Manager, VPC Service Controls and the Wiz integration stack.

Prepared by the Softprom editorial team based on vendor materials (Softprom DistriFlow). Source: original article.