How do I stop an AI agent with cloud credentials from running up a huge bill by launching GPU instances?

What should I check, step by step?

1. Give the agent its own AWS account, Google Cloud project or Azure subscription, so its spend, quotas and budgets are separate from production and can be cut off without touching anything else.

2. Grant a role scoped to the task, not an admin key. On AWS, deny ec2:RunInstances unless ec2:InstanceType is on an allowlist (AWS's EC2 policy examples show this pattern for t2.micro and t2.small). Put the same deny in a service control policy so the agent can't attach a broader policy to itself.
3. Keep GPU and accelerator quotas at zero, or as small as the task needs, in the agent's account. On AWS the default On-Demand vCPU quota for P, G and other accelerated families is 0, but AWS can raise quotas automatically based on usage, so check them and never let the agent's role request increases.
4. Put a gate in front of every create call that costs money: estimate hourly cost times maximum runtime, and require a human approval above your threshold before the API call runs. Deny instance types or regions you didn't plan for.
5. Set a maximum lifetime on everything the agent launches: tag it with the run id and an expiry time, and run a scheduled job that stops or terminates anything past its expiry. Delete orphaned volumes and snapshots too.
6. Add budget actions as a backstop, not the main control. AWS Budgets actions can apply an IAM deny policy or SCP, or stop specific EC2 or RDS instances, automatically or after approval. Google Cloud can disable billing on a project from a budget Pub/Sub notification, which shuts down all its services.
7. Use Google Cloud spend cap budgets where they apply (Gemini API, Vertex AI, Cloud Run, Cloud Run functions, one project and one service per budget). Compute Engine VMs aren't covered, and Google says enforcement isn't instant.
8. Set budgets below the money you can actually afford, because reporting lags. Google recommends a budget below your available funds for this reason, and Azure says cost data is typically available within 8–24 hours.
9. Watch for auto-recovery: AWS warns that if a budget action stops an instance in an Auto Scaling group, Auto Scaling restarts it or launches a replacement. Cap the group's maximum size too.

Why don't AWS, Google Cloud or Azure budget alerts stop an AI agent's runaway cloud spend?

An agent can launch a fleet of GPU instances in seconds, and they bill by the second or hour while the bill catches up. Cloud budgets are built on billing data that arrives later: AWS updates the data Budgets uses at least once a day, Google Cloud says there is a delay between use and costs reaching Cloud Billing, and Azure says cost data typically shows up within 8–24 hours.

• Google Cloud's documentation says alerts-only budgets don't automatically prevent the use or billing of your services, and Azure's says resources aren't affected and consumption isn't stopped. Even Google's own sample that disables billing from a budget alert warns that it doesn't guarantee you won't spend more than your budget.

How do I stop an AI agent from launching expensive GPU instances in the first place?

Permissions are checked when the API call is made, with no billing lag. An IAM or SCP deny on unapproved instance types, a zero GPU quota and a separate account for the agent stop the expensive launch from happening at all.

• For launches you do allow, an approval gate turns 'the agent decided' into 'a person agreed to this instance type, count, region and maximum runtime'. Approve the exact request (instance type, count, region, maximum hours and estimated cost), not a summary of it, and refuse the call if what is about to run differs from what was approved.

What do AWS Budgets actions and Google Cloud spend caps do when an agent overspends?

AWS Budgets actions can run automatically or after your approval when a threshold is crossed: apply a deny IAM policy or SCP so nothing new is provisioned, or target specific EC2 and RDS instances. From the management account you can apply an SCP to another account, but you can't target instances in another account.

• Google Cloud spend cap budgets pause new usage of an eligible service in one project when estimated gross costs pass the budget, until you lift the cap or the month resets. In-flight requests finish and are billed, persistent resources such as compute and storage keep accruing, and overages are billed as normal.
• Treat these as a second line. They cut off spend after the fact, sometimes along with services you wanted to keep running.