Google Cloud used its Security Summit on November 4, 2026 to ship a set of controls for companies that now let AI agents write code and call internal systems. The announcements, outlined by Sandra Joyce, vice president of Google Threat Intelligence, cover three areas: checking code that agents produce, limiting what deployed agents can reach, and automating parts of the security operations center.
The most concrete change is to CodeMender, Google's autonomous code-security agent, which has been in preview since July. According to Google's documentation, CodeMender wraps a language model in a harness built with Google DeepMind prompts, skills and orchestration, then finds a vulnerability, tries to exploit it to confirm it is real, and proposes a tested patch. The new version runs that verification and exploit simulation inside an operating-system sandbox or an isolated virtual machine chosen by the customer, and asks a person to approve local file writes and command execution. Teams can also limit a scan to one folder, a whole repository, or only the files changed since the previous scan, which Google presents as a way to control speed and cost.
Google says hundreds of customers have used the preview. It cites Wiz, the security company it owns, using Gemini 4 Argon alongside other models to find and help fix the core PHP flaws disclosed in October, and quotes Thales saying its early evaluation points to a possible cut of more than 95% in vulnerability discovery time. Those figures come from Google and its customers; no independent measurement was published. The product documentation also notes that CodeMender is still a pre-GA offering for a limited set of customers and is not meant for production use under its preview terms.
The second group of releases targets agents once they are deployed. Auth Manager, part of Google's Agent Identity service, is now generally available. It keeps credentials in a Google-managed vault governed by IAM, runs OAuth 2.0 and PKCE flows, and refreshes user-delegated tokens so agents can act on a person's behalf without long-lived secrets. In preview, Agent Identity can also issue standard OIDC tokens in JWT form, letting an agent authenticate to AWS, Azure, on-premises services and MCP servers without static API keys.
Other additions are smaller. VPC Service Controls rules can now be written around IAM roles, with Workload Identity Federation for GKE in preview. Model Armor gains a custom topic detector in preview that flags conversations drifting off an approved subject, aimed at regulated sectors such as healthcare and insurance. Cloud NGFW adds a malware sandbox based on Palo Alto Networks WildFire, and Cloud Armor gets managed rules in preview backed by Thales Imperva threat data.
Wiz features heavily in the package. Google announced Wiz AI SAST in preview, a static analysis tool that uses agent-style reasoning to look for business logic flaws and hands fixes to CodeMender, plus a shared data model between Wiz Defend and Google Security Operations. Scan for Good, a joint Google DeepMind and Wiz effort, has helped more than 50 public service and critical infrastructure organizations find exposures in its first month, Google says.
The summit follows Google's October 22 launch of Agent Compute Engine, its microVM service for agents, and fills in the identity and policy side of that stack. Google did not publish prices for the new features or a date for CodeMender to leave preview.