Google Cloud adds identity and code-fixing tools to secure AI agents

What Google released

As IT Brief Asia reported, Google Cloud has rolled out a set of security products and updates for organisations that use autonomous AI agents, meaning software that can take actions on its own rather than just answer questions.

The package works on three fronts: protecting the way software is written, locking down agents once they run in production, and connecting threat detection and response between Google's security products and Wiz, the cloud security company whose platform Google now works closely with.

Some items are generally available, meaning open to all customers. Others are in preview, meaning early access that may still change.

CodeMender: finding bugs and testing fixes

CodeMender is Google Cloud's code security agent, and it is still in preview. Its job is to confirm that a suspected vulnerability is real, then produce a patch that has already been tested, so teams are not stuck with a long list of unverified warnings.

The update adds isolated checking: CodeMender can simulate an exploit inside either an operating-system sandbox (a walled-off area) or a virtual machine. Humans still have to approve steps that write files locally or run commands. Customers can also choose how much to scan: one folder, a whole repository, or just the files changed since the previous scan.

Google says hundreds of customers have tried it since the preview began in July. It also pointed to work by Wiz using Gemini 4 Argon alongside other models on a flaw in core PHP, which Google says affects 70% of websites. According to Google, the model helped locate the bug, map a likely attack route, detect attacks and confirm the fix worked.

Thales Group's Ivan Fontarensky said an early evaluation suggests CodeMender could cut vulnerability discovery time by over 95%, with experts still validating the results. That is an early customer view, not a measured result. IDC's Katie Norton put the underlying problem plainly: nearly 40% of enterprises add vulnerabilities faster than they can fix them, and, in her words, "Detection without remediation only adds to developer workloads."

Identity: agents without permanent keys

Google says attackers increasingly go after the credentials, data boundaries and connections that tie agents to company systems. Several of the new controls address that.

Auth Manager is now generally available. It stores credentials in a vault run by Google and governs them with the customer's identity and access management rules. It handles the OAuth 2.0 and PKCE sign-in steps (standard methods for letting software act for a user) and renews the tokens it holds on a user's behalf. Google says this means fewer repeated consent prompts and less chance of stolen credentials.

Agent Identity JWT support is in preview. It gives agents standard OIDC tokens, a common format for proving identity. Agents can use these to sign in to Amazon Web Services, Microsoft Azure, on-premises systems and MCP servers (the connectors agents use to reach tools and data) without a static, long-lived API key that could leak and be reused.

Google is also tying identity to network limits. Rules for VPC Service Controls, which set what data may enter or leave a protected cloud perimeter, can now reference IAM roles, and that feature is generally available. Workload Identity Federation for Google Kubernetes Engine is in preview. Google says the aim is finer perimeter control across enterprise clusters without piles of hand-written exceptions.

Inspecting what agents say and send

Model Armour has a new custom topic detector, in preview. Organisations write rules in plain prompts to check whether an agent's conversations stay inside approved subjects. Google names healthcare, insurance and retail as target sectors.

On the network side, Cloud NGFW's advanced malware sandboxing is generally available and runs on Palo Alto Networks WildFire. Cloud NGFW project-level resources are also generally available. Cloud Armour managed rules are in preview, supported by threat intelligence from Thales Imperva.

Deeper ties to Wiz

Wiz Defend and Google Security Operations now use a shared data model and synchronise in both directions. The link adds a Wiz Content Pack, Wiz Blue Agent investigation analysis inside Google Security Operations, and a connection between Wiz Defend and Google Threat Intelligence.

Wiz AI SAST is also in preview. SAST means static application security testing, which examines code without running it. This version uses AI reasoning to study how an application behaves and spot business-logic flaws, where the code works as written but allows something it should not. Findings are matched against infrastructure context in the Wiz Security Graph.

Separately, Google DeepMind and Wiz have started a joint effort called Scan for Good with public service, critical infrastructure and non-profit groups. Google says Argon helped more than 50 such organisations find and fix critical exposures in the first month.

How to read the launch

The pattern across these tools is consistent: give each agent its own short-lived identity, limit what it can reach, keep a human in the loop for risky actions, and test fixes before shipping them. Most figures here come from Google and early customers, so treat them as claims to test in your own environment.