Discussion

Google Cloud adds sandboxing and tighter controls to AI security tools

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#11471

Google Cloud is adding ways to constrain and manage AI security tools, including isolated exploit simulations, human checks on sensitive actions and more targeted code scans. The changes give security teams more control over where agents operate and what they can touch, while CodeMender remains in preview.

Watch Desk analysis

What happened

The updates cover both CodeMender, Google’s AI tool for finding and fixing software vulnerabilities, and controls for agents running across Google Cloud services. Help Net Security reports that CodeMender can verify vulnerabilities and simulate exploits inside an operating-system-level sandbox or isolated virtual machine. Teams can scan a whole repository, a folder or only files changed since the last scan.

The tool also adds human checks for local writes and command execution. Google Cloud’s wider agent controls include Auth Manager, now generally available, which centralises credentials in a Google-managed vault. Agent Identity JWT support, which lets agents authenticate across services without static, long-lived API keys, is in preview. A custom topic detector for Model Armor is also in preview. (Help Net Security’s report)

Why it matters

Security agents need enough access to inspect code and investigate threats, but that access can itself create risk. Sandboxed testing, narrower scans and checks before local writes or command execution are practical guardrails, not merely another promise that an agent will behave.

The controls also matter beyond CodeMender. Centralised credentials and agent identity could help organisations manage how agents reach other services, while topic rules offer a way to flag conversations outside approved boundaries. Several capabilities remain in preview, so availability and operational performance will vary.

Our read

This is a useful shift from asking whether AI can help with security to asking how it should be allowed to act. The best detail is the human check on actions that change local files or run commands: autonomy is more useful when there is a clear brake pedal.

For now, teams should distinguish generally available controls from preview features and test the latter against their own workflows. A tidy control plane is welcome; it still needs to work under real security workloads.

What to watch

  • Whether CodeMender moves beyond preview, and what safeguards remain in place.
  • Whether organisations adopt agent identity instead of long-lived API keys.
  • How teams configure human approval for local writes and command execution.
  • Whether Google publishes measurable results from real-world deployments.

Discussion spark: For AI security agents, which actions should always require human approval, even if that makes investigations slower?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.