Google adds agent security controls and new CodeMender scan modes

At the Google Cloud Security Summit, Google announced security tools for AI agents, which are software that operates with no person. CodeMender, its agent that finds and repairs code bugs, is in preview, and Google tells that hundreds of customers used it since July. Auth Manager, a vault for the logins of agents, is available to all customers.

What is new

• CodeMender makes attack tests in a sandbox or an isolated virtual machine.
• A person must examine each write of local files and each command.
• It can examine one folder, one repository or only the files that changed.
• Model Armor topic rules keep agent conversations on topic, in preview.

The numbers

Thales, a partner, tells that a first test shows CodeMender can cut the time to find bugs by more than 95%. Google tells that its model helped find core PHP bugs that affected 70% of websites. The firm IDC tells that almost 40% of enterprises add bugs but cannot repair them as fast as they add them.

What is not known

The number from Thales is a first view from a partner. Google gives no test data. CodeMender is in preview, and Google gives no price.

Insight: Google supplies the agent, its cage and the guards together

Google gives each agent a small virtual computer in Agent Compute Engine. At this time it adds identity, network and content controls, with tools from partners. This reads as a race to own the gate for agent work, because the cloud that operates agents can also set its security rules. One gate is easy to get but hard to change.