Skip to content
Common question

How do I stop an AI agent from burning my API key?

You stop an AI agent from burning an API key by isolating where the key lives, limiting which agent can use it, and running the agent on a controlled cloud machine, not by pasting the same key into every local script and shared host.

Get startedRead the docs

“Burning” usually means runaway loops, retries, tool spam, or one compromised process using a key that also powers production. The fix is operational: scope, isolation, and a host you can shut down or rotate without guessing which laptop still has the secret.

Common ways keys get burned

  • The same provider key is reused across demos, clients, and production agents.
  • The agent runs on a shared machine or laptop where other processes can read the env.
  • There is no clear per-project boundary, so a buggy agent keeps calling the API until the bill or rate limit hits.

Practical controls that actually help

  1. One key (or credential set) per project/agent where the vendor allows it. Do not share production keys with experiments.
  2. Isolate the runtime so shell access in one agent does not mean free access to every .env on a shared host.
  3. Prefer a dedicated cloud box per project so you can stop, rebuild, or rotate without tearing down unrelated workloads.
  4. Use connectors and channels deliberately. Only attach what that agent needs (email, GitHub, Notion, and so on), not every secret you have.

None of this replaces provider-side spend limits and monitoring. Those still matter. Isolation and scoping reduce how bad a single agent failure can get.

How Cloudrun fits

Cloudrun deploys each AI agent to its own isolated cloud machine, with a single fleet canvas. Per-project isolation and a baked runtime snapshot support keeping credentials and connectors with that agent’s box instead of on a shared laptop or shared host.

That same separation is why agencies use per-client boxes. See How do agencies host separate AI agents per client without shared blast radius?.

Related

Pillar: Deploy an AI agent to an isolated cloud machine.

Next step

Run agents with per-project isolation at usecloudrun.com.

Keep reading

Related Cloudrun guides

Direct answers

Questions this page answers

How do I stop an AI agent from burning my API key?
You stop an AI agent from burning an API key by isolating where the key lives, limiting which agent can use it, and running the agent on a controlled cloud machine, not by pasting the same key into every local script and shared host.
Common ways keys get burned
The same provider key is reused across demos, clients, and production agents. The agent runs on a shared machine or laptop where other processes can read the env. There is no clear per-project boundary, so a buggy agent keeps calling the API until the bill or rate limit hits.
Practical controls that actually help
Use one key (or credential set) per project or agent where the vendor allows it. Isolate the runtime so shell access in one agent does not mean free access to every .env on a shared host. Prefer a dedicated cloud box per project so you can stop, rebuild, or rotate without tearing down unrelated workloads. Attach only the connectors and channels that agent needs.
How Cloudrun fits
Cloudrun deploys each AI agent to its own isolated cloud machine, with a single fleet canvas. Per-project isolation and a baked runtime snapshot support keeping credentials and connectors with that agent’s box instead of on a shared laptop or shared host.
cloudrun

Keep keys on the project box

Run agents with per-project isolation at usecloudrun.com.

Get startedView the product