Google rolls out new tools to guard AI agents and the code they write

Google has announced new security tools for companies that use AI to write code and run tasks. The best known is CodeMender, an AI agent that finds and fixes flaws in software code. It has been in preview since July, and Google says hundreds of customers have used it. The news came at the Google Cloud Security Summit.

What changed

Google says AI coding tools have sped up software delivery so much that human review cannot keep up. It also warns that agents can become a back door if they are trusted blindly. CodeMender, now in preview, gets two additions:

• Isolation: flaw checks and exploit tests run inside a sandbox or an isolated virtual machine, a sealed space that keeps tests away from real systems.
• Human checks: a person must approve local file writes and commands.
• Scan options: scan one folder, a whole repository, or only files changed since the last scan.

What the early results say

Google points to Wiz, a security company. Wiz used Gemini 4 Argon, a Google AI model, along with other models to help find and fix core PHP flaws that affected 70% of websites. Thales, a cybersecurity firm, says its early tests show CodeMender could cut flaw-finding time by over 95%, with experts still validating fixes. These are early claims from Google and its partners.

Analyst firm IDC adds context. It says nearly 40% of enterprises add vulnerabilities faster than they can fix them.

Guardrails for AI agents

Google says attackers go after the plumbing that connects agents to company systems, such as credentials and data boundaries. Its answer is a set of controls for agents on Google Cloud and across other clouds. Some are generally available, and some are still in preview:

• Auth Manager, generally available: keeps agent credentials in a Google-managed vault.
• Agent Identity JWT support, in preview: lets agents sign in elsewhere without long-lived keys.
• Model Armor custom topic detector, in preview: keeps agents on topic.
• Cloud NGFW advanced malware sandbox, generally available: enabled by Palo Alto Networks.

What to do

Most small businesses will not use these tools directly. The lesson still applies. If your team uses AI tools to write code or run tasks, check what access those tools have. Ask who reviews their work before it goes live. Google says over-privileged credentials are a common way attackers spread once inside.