Independent: no vendor pays for coverage · Built from vendor documentation and Google's threat research · Prepared with AI assistance; documentation review · How we test · Edited by Annie Vane
If your business uses an AI API key (OpenAI, Anthropic, Google), switch on a hard monthly spend limit today, not just an alert. Alerts email you; only a hard limit (or, on Google Cloud, a spend cap or automation) stops the bill. Google's threat researchers reported on 2 November 2026 that criminals are stealing AI API keys and cloud credentials and reselling access through "LLM mixers", running up charges on victims' accounts. A hard limit, separate keys per project and keeping keys out of code won't stop a theft, but they cap how much it can cost you. Each one is a few settings in the vendor's billing pages, according to their documentation.
Last checked: 8 November 2026 · Written by Annie Vane, editor · Evidence: Documentation review (vendor docs + Google Threat Intelligence report)
Prepared with AI assistance; documentation review. No hands-on test. I did not test these settings on a live account; menu names are from each vendor's documentation and may have moved. How we test
What's happening: "LLMJacking" and LLM mixers
Google Threat Intelligence Group (GTIG) and Mandiant published a report on 2 November 2026 on what they call LLMJacking: attackers hijacking a company's access to AI models and running up charges on its account. According to the report:
- The main way in was exposed or stolen cloud credentials. Attackers also harvested AI API keys from public GitHub repositories.
- Attackers tested stolen keys with tiny requests (one output token) so they wouldn't trigger volume alerts.
- Access was then pooled and resold through "LLM mixers", illicit gateways that proxy stolen access to paying subscribers.
- GTIG says it saw fraudulent charges of more than US$100,000 in a single day in some incidents, and cumulative charges reaching about US$1 million at some organisations before access was revoked.
Those figures are Google's own incident findings. We haven't verified them, and the incidents it describes were at organisations in North America, South America and Asia Pacific. CSO Online reported the research on 3 November 2026. For a small UK firm the lesson is simpler: if a key leaks, assume the bill lands on you unless the provider agrees otherwise. OpenAI's and Anthropic's terms make the customer responsible for activity under its account. OpenAI's services agreement says fees are non-refundable except as required by law or as the agreement permits, and Anthropic's credit terms say its credits are non-refundable.
Alerts warn, limits stop

All three offer alerts. OpenAI and Anthropic also offer hard spend limits; Google Cloud offers a spend cap on some services, or automation you set up yourself. Their documentation is clear that alerts alone don't stop spending:
| Control | What happens when you hit it | Good for |
|---|---|---|
| Spend alert / budget alert | Sends an email. Traffic carries on. | Early warning |
| Hard spend limit | API requests start failing. Spending stops (after a short lag). | Capping the damage |
A hard limit can break your own app when it's reached. That's the trade-off. Set it above normal usage with headroom, and put an alert at 50 to 80 per cent of it so you hear first.
How to set a hard limit, vendor by vendor
Settings change. These steps are from each vendor's documentation as of 8 November 2026; check your own console.
OpenAI (API platform)
OpenAI supports monthly spend limits at organisation and project level.
- Go to Project settings and select Limits. OpenAI's older help article describes reaching this from organisation settings, so confirm the menu names on the day.
- Under Spend, choose Edit spend limit and enter the Monthly spend limit.
- Turn on Enforce a hard limit. OpenAI says this is what makes API responses fail once the limit is reached; spend alerts on their own do not enforce a cap.
- Add alerts at lower thresholds so you hear before the limit bites.
OpenAI says enforcement "is not instantaneous", so recorded spend can slightly exceed the limit. Once reached, requests return a 429 error (organization_spend_limit_exceeded or project_spend_limit_exceeded). This is separate from the usage limit OpenAI itself assigns by usage tier.
Anthropic (Claude Console)
- Organisation: on the Billing page, in the Spend limits section, choose Adjust limit and set your own limit. It can't exceed your tier's cap.
- Per workspace: on the Workspaces page, select a workspace and open Spend limits to cap it and set alerts. Anthropic says you can't set spend limits on the Default Workspace, so put real work in a named workspace.
When you reach a limit you set, Anthropic says the API returns an HTTP 400 invalid_request_error saying you have reached your specified API usage limits. That is different from hitting your tier's cap, which returns a 429.
Google Cloud (Gemini API and Gemini Enterprise Agent Platform, formerly Vertex AI)
Google's documentation is blunt: an alerts-only budget "doesn't automatically cap" spending. Your options:
- Spend cap budget on a project, for eligible services only. Google's docs (checked 8 November 2026) list Gemini API and Gemini Enterprise Agent Platform (formerly Vertex AI), plus Cloud Run and Cloud Run functions. When enforced, all new usage of that service in that project pauses for the rest of the month, until the cap lifts the next month or you lift it manually. Requests already running finish and are charged, and persistent resources keep billing. Alerts-only budgets still do not stop spend.
- Otherwise, an alerts-only budget plus a Pub/Sub notification that triggers automation, such as disabling billing on the project. That's a developer job, and Google notes cost reporting can lag.
Google also suggests setting a budget below your available funds to allow for reporting delays.
Another layer: restrict where a Claude key works
A network restriction is not a spend cap. Anthropic's 5 November release notes introduce an organisation IP allowlist in beta, with up to 1,000 IPv4/IPv6 addresses or CIDR ranges. Requests to covered endpoints from unlisted public IPs are refused. Coverage is limited: consult the allowlist documentation for the covered Console/API services and exclusions; do not assume it controls every Claude product, partner platform or managed-agent runtime connection. It can narrow where a leaked key works, but it does not set a budget, shorten a key's lifetime or replace stronger identity controls. Use key expiry or short-lived workload identity where appropriate as well as spend controls, and plan around lockout and changing network addresses. Evidence: documentation review, 8 November 2026 (Europe/London), not a console test.
Five key habits that cost little or nothing
- One key per app and per environment. Separate keys (and OpenAI projects or Claude workspaces) for live, testing and each staff tool. A leak then affects one budget, and you can revoke one key without taking everything down.
- Least privilege. Give each key only the models and permissions it needs. OpenAI lets you set a project key to Restricted scopes and limit which models a project can use; Google Cloud lets you restrict a key to specific APIs plus one type of application restriction (IP addresses, websites, or Android or iOS apps), and AI Studio offers a restriction to the Gemini API only. Restrict Google API keys to the specific APIs they call.
- Never put keys in code, documents or chats. Use environment variables or a secrets manager. Turn on secret scanning in GitHub so a committed key is caught. GTIG specifically mentions keys harvested from public repositories.
- Rotate on any doubt. If a key may have leaked (a public repo, a departed contractor, a shared laptop), revoke it and issue a new one first, then investigate. Check usage for spikes and keys you didn't create: GTIG says attackers created extra keys and accounts once inside.
- Watch for odd usage. A sudden jump in requests, models you don't use, or activity at 3am are the warning signs. Tiny probe requests are designed to stay under the radar, so a spend limit is your backstop.
Keys used by AI agents and add-on "skills" deserve extra care. See our AI agent security notes.
Ask your supplier the same question
If you buy an AI tool rather than holding keys yourself, the supplier holds the keys. Question 8 of our supplier due-diligence checklist asks what it costs when usage spikes and whether there's a hard cap. A supplier billing you in "credits" with no cap may be exposed to the same risk, and so may you.
If the supplier's tool uses Claude, ask whether it supports Anthropic's linked billing, so you don't have to hand over a key at all. Our guide to stopping pasting API keys into AI tools explains how it works.
Frequently asked questions
Does a spend alert stop charges?
No. OpenAI and Google both say alerts on their own don't cap spending. To stop traffic you need a hard spend limit (OpenAI, Anthropic) or a spend cap budget or automation (Google Cloud).
Can spend go over a hard limit?
Slightly. OpenAI says enforcement isn't instantaneous, so recorded spend can exceed the limit a little. Google says cost reporting is delayed. Set limits with that in mind.
What should I do first if I think a key has leaked?
Revoke the key and create a new one. Then check usage and billing for spikes, look for keys or service accounts you didn't create, and contact the vendor's support about the charges.
Is ChatGPT Plus or Claude Pro affected?
These are flat-rate subscriptions, not pay-as-you-go API keys, so this kind of runaway billing is mainly an API-key risk. Both companies do sell optional extras: Anthropic's paid Claude plans can turn on prepaid usage credits for use beyond plan limits, billed at standard API rates, with a monthly spend limit you can set; OpenAI sells pay-as-you-go credits for some ChatGPT features, such as Codex. Claude's usage credits are off until you enable them, and OpenAI's credits have to be bought, so check your settings if you have.
Sources
- Google Threat Intelligence Group, The rise of LLMJacking: how threat actors monetize compromised cloud and model resources, 2 November 2026 (vendor research; figures are Google's)
- Secondary: CSO Online, Enterprises pay $100,000 a day for AI they never ordered (3 November 2026)
- OpenAI, Spend limits and Managing projects in the API platform
- Anthropic, IP allowlisting and release notes, 5 November 2026, read 8 November 2026
- Anthropic, Rate limits (spend limits section) and Workspaces
- Google Cloud, Create, edit, or delete budgets and budget alerts and spend cap budgets
- Google Cloud, Add restrictions to API keys
- OpenAI, Services Agreement; Anthropic, Commercial Terms and Supplemental Credits Terms
- Anthropic, Manage usage credits for paid Claude plans
This page explains vendor settings and published security research in plain English, checked against each vendor's documentation on 8 November 2026. Settings, menus and limits change, so verify them in your own console. It is general information, not security, legal or financial advice. If you think a key has been abused, contact the vendor and your IT or security provider straight away. Spotted something out of date? Tell me.



