Five Bearings
Who Pays When Your AI API Key Is Stolen? Often, You Do
Stolen AI keys ran up US$740,000 on one platform's customers in October. Refunds are up to the provider. Five controls Australian SMEs can set this week.
In short. Usually the account that owns the key gets the bill, and any refund is at the AI provider's discretion. After stolen keys ran up about US$740,000 on Railway's customers in October 2026, AWS and Google began refunds, OpenAI reviewed claims case by case and OpenRouter refused. Insurance may not respond either. The fix is a hard spending cap, one scoped key per tool and keys kept out of code.
Does software help decide things about your customers or staff? From 10 December 2026 your privacy policy may have to say so. Six questions give a likely answer.
Take the check →One exposed key, a six-figure bill by morning
In October, attackers who broke into the app-hosting platform Railway used its customers' stored AI keys to run up about US$740,000 in AI usage, roughly A$1.07 million according to iTnews. Google's threat intelligence team says victims of this kind of theft have been charged more than US$100,000 in a single day. The bill lands on whoever owns the key. Whether any of it comes back is the AI provider's call.
It isn't only big platforms. In April an Australian AI consultant woke to a Google Cloud bill of A$25,672.86 on an account with a A$10 budget, as reported by Tom's Hardware. The budget was an alert, not a limit. By the time it fired, A$10,000 had already gone through on his credit card.
US$740k
Stolen AI usage on Railway customers' keys
October 2026, per Railway
US$100k+
Charged to some victims in one day
Google Threat Intelligence, Nov 2026
A$25,672
One Australian's overnight bill
On a A$10 budget alert
What happened in October
Security firm Wiz counted at least nine AI developer platforms that disclosed incidents in October 2026, Railway among them. Railway's incident report says the attacker got in through a flaw in its build infrastructure on 20 October and was evicted the next day, having decrypted stored settings for up to 1.85% of workspaces and searched specifically for ones named like AI provider keys.
Google calls the practice LLMjacking: thieves run their own AI workloads on your account, or resell the access cheaply, and you pay. In one case Google describes, a reseller pushed nearly US$1 million of usage onto a single victim in under eight hours.
Then came the refunds, and they were uneven. Per Railway, AWS and Google began refunding under their fraud policies, OpenAI reviewed claims case by case, Anthropic was still in discussion, and the gateway OpenRouter enforced its no-refund terms. Railway said it would cover whatever providers didn't. That was the host's decision. A small business can't plan on a stranger's goodwill.
Why this is an Australian SME problem, not a Silicon Valley one
You may not think you have an API key. But if an agency built your website chatbot, if ChatGPT, Claude or Gemini is wired into Zapier or Make, if a developer built you a quoting or document tool, or if someone published an app from Google AI Studio, there is a key somewhere, and it is billing a card.
The Australian Signals Directorate warned in late September, in its advisory Protect your organisation's AI services, that attackers are reaching AI services through compromised credentials, authentication tokens, user sessions, vulnerable applications and third-party access arrangements. Its message to businesses was to treat AI access as security-sensitive rather than rely on the AI developer's controls.
Don't assume insurance catches it. Insurance Business reported in September that most Australian cyber wordings carry no AI-specific language, so whether a stolen-key bill is covered gets decided at claims time. It cited a Gallagher survey in which one in five insurance professionals said a client had suffered AI-related losses in the past year, and just over half of those losses were fully covered.
This is a cost intelligence problem as much as a security one. Your AI spend is only predictable while the only person using your key is you. As we wrote when covering what happens when an AI vendor fails, the risk tends to stay with the business that signed up. If you don't know how many keys your business has or who created them, we can map them with you in a single call.
Five controls you can set this week
1. Inventory. List every AI account and key: who created it, which card it bills, which tool uses it, where it's stored. The ASD advisory recommends exactly this. Include keys an agency or contractor set up on your behalf.
2. Hard caps, not alerts. An alert tells you about the damage. A cap stops it. Set a monthly usage limit with each provider and, where prepaid credit is offered, use it and switch off auto top-up. Google's guidance for defenders lists spending caps and per-model quotas alongside budget alerts, not instead of them. We made the case for capping AI spend earlier this year; this is the security reason.
3. One key per tool, scoped tight. Give the chatbot, the automation and the internal tool separate keys, each limited to the models it needs. Google recommends this so that a stolen key for a cheap model can't call the expensive ones, and so you can kill one key without breaking everything.
4. Keys out of code, spreadsheets and shared docs. Keep them in a password manager or a proper secrets store. Rotate them when a developer, agency or staff member who had access moves on.
5. Rehearse the bad morning. Know how to revoke a key in each provider's console and who to contact about disputed charges. Then ask your broker one question at renewal: does our policy respond to unauthorised AI usage charges?
None of this is a reason to keep AI out of the business. It's the same discipline you already apply to the company card, applied to a card that thieves can now spend at machine speed.
Key takeaways
Common questions
Who pays if my AI API key is stolen?
Usually the account that owns the key is billed, and any refund is up to the AI provider. After the October 2026 Railway breach, AWS and Google began refunds under fraud policies, OpenAI reviewed claims case by case and OpenRouter enforced a no-refund policy.
Does a budget alert stop AI charges?
No. An alert notifies you; it doesn't stop usage. An Australian AI consultant was billed A$25,672.86 overnight on a Google Cloud account with a A$10 budget, and A$10,000 had been charged before the alert arrived. Set hard usage caps or prepaid limits as well.
Does cyber insurance cover stolen AI key charges in Australia?
Not reliably. Insurance Business reported in September 2026 that most Australian cyber policy wordings have no AI-specific language, so coverage is decided at claims time. Ask your broker directly whether your policy responds to unauthorised AI usage charges.
Sources
Assumptions & methodology
- Last reviewed 14 November 2026 · general information for Australian businesses, not legal, insurance or security advice.
- Railway figures (US$740,000 estimated unauthorised spend, up to 1.85% of workspaces, 20–21 October timeline, provider refund positions) are from Railway's incident report. The A$1.07 million conversion is iTnews's, from its report of 11 November 2026. Refund positions are as Railway described them at the time and may have changed.
- The 'at least nine platforms' figure is from Wiz's 'Black October for AI developer platforms' post (30 October 2026, updated to 10 November). The US$100,000-a-day and nearly US$1 million figures are from Google Threat Intelligence Group's LLMjacking report of 2 November 2026, which also sets out the defensive controls referred to here.
- The A$25,672.86 bill (Jesse Davies, an Australian AI consultant) is as reported by Tom's Hardware on 22 April 2026, citing his own posts. His key was exposed through a publicly reachable service rather than stolen in a breach, and the charges were reportedly waived.
- The ASD advisory page could not be loaded when this note was written; the summary of its warning and recommendations is as reported by Insurance Business (29 September 2026) and other coverage. The Gallagher survey figures are as reported by Insurance Business.
- The examples of where a small business's keys may sit (chatbots, Zapier or Make, custom tools, AI Studio apps) are our illustrations.
Talk to us
Do you know how many AI keys your business has?
We help Australian businesses run AI with the same controls they'd put on a company card: who holds the keys, what each one can spend, and what happens when one leaks. Book a call and we'll go through yours.