ML in Production · 4 min read
Stolen AI API Keys Now Cost Victims Up to $1 Million
Google and Mandiant report LLMJacking losses up to $1M; WardenStealer has logged 130,000+ infections since August. Stolen AI keys are now an industrial trade.
Key takeaways
- Google and Mandiant (Nov 2026): some orgs lost over $100,000 in fraudulent charges in one day; totals reached $1 million.
- SpyCloud logged 130,000+ unique WardenStealer infections since August 2026, targeting Claude Code, Codex, Cursor.
- Black-market prices for some AI coding accounts more than doubled during 2026, per Google.
- Amazon researchers documented AI agents harvesting keys from exposed LiteLLM, Dify, RAGFlow, Open WebUI deployments.

Organizations hit by LLMJacking have lost more than USD 100,000 in a single day, and in some cases total losses reached USD 1 million before anyone shut the abuse down. Those figures come from Google and Mandiant's November 2026 report, and they describe the outer edge of a market that has quietly industrialized: stolen AI API keys, compromised accounts, and discounted proxy access to commercial models, all traded underground.
The mechanics are unglamorous. An API key is a digital pass that lets an application call an AI provider's models. Developers stash these keys in code files, cloud systems, local devices, and AI applications. One exposed key hands criminals paid services without the owner's consent. Criminals phish for them, harvest them from exposed code repositories and weak security settings, or simply steal them with malware. Once they test what a key unlocks, it joins a larger pool that other criminals buy or use.
How big is the underground market?
Google's Threat Intelligence Group reported in November 2026 that criminal groups have built a genuine business around stolen AI access. Some operators gather compromised credentials; others handle sales and customers. Google also described a group called LLM Mixers, which operates as an illegal gateway to commercial AI models.
The catalog runs well past single keys. Sellers trade compromised AI accounts, token quotas, cloud access, and proxy services that let buyers run commercial models at a discount. Access to popular AI coding tools and paid subscriptions is on the shelf too.
Prices tell the demand story. According to Google, black-market prices for some AI coding subscriptions and accounts more than doubled during 2026. That's roughly the appreciation of a hot growth stock, applied to someone else's login.
What malware is targeting developers?
SpyCloud Labs reported on October 30, 2026 that it had logged more than 130,000 unique WardenStealer infections since August. The malware goes after credentials and configuration data tied to Claude Code, Codex, and Cursor. A second family, Aoki, also hunts AI-related credentials, combing devices for stored secrets, account details, and session data.
Amazon researchers described a related campaign in which AI agents searched for exposed AI applications — LiteLLM, Dify, RAGFlow, and Open WebUI. The agents looked for provider keys and dropped valid ones into shared pools. Some records suggested the stolen keys were then used to find even more exposed credentials. A self-feeding loop, in other words.
Why is LLMJacking hard to catch?
LLMJacking covers the unauthorized use of someone else's AI model access or compute. Criminals send requests through the victim's account and leave the legitimate owner with the bill. Detection is the problem. A sudden jump in AI usage can look like ordinary business activity, especially at companies that lean on AI daily, and proxies plus multiple compromised accounts obscure where requests originate.
The attack surface also extends past sloppy repositories. Anthropic's September 2026 threat report described fake AI websites and apps advertising discounted access to commercial models — a few collected account details from customers chasing a bargain. Malicious coding apps can lift login details and session tokens straight from developer devices, so people who never run exposed servers or publish code are still exposed. A cheap AI offer can cost you a much larger set of personal or business accounts.
What should data teams actually do?
The defensive playbook is boring, which is the point:
- Treat AI API keys as sensitive business secrets. They belong in secure secret managers, not source code or shared files.
- Use separate keys for test and production systems, which limits the damage when one leaks.
- Set usage limits, review API activity, and alert staff when token use climbs without a clear reason.
- Audit cloud permissions and protect AI gateway settings.
- Revoke any exposed credential immediately — and note that resetting a key may not suffice if criminals also hold cloud logins or active session tokens.
Does this matter beyond developers?
Yes. The market for stolen AI keys shows how easily access to paid technology converts into income. One compromised account feeds illegal sales, enormous AI bills, and further cyberattacks, and the risk now stretches well beyond software developers. As more companies depend on AI tools, weak credential security invites financial losses and data exposure across the wider digital economy.
The real challenge is control: who can use AI resources, where credentials live, and what every request costs its owner. A company that fails to protect its AI access may end up paying for the very tools criminals turn against it.
spycloud.com (Original)
More from Grace Kim
Show full bio
Market editor covering companies, deals and strategy in Analytics Digest is a professional daily at Analytics Digest.
22 articles