Skip to content

LLMjacking: How Attackers Steal Cloud AI Credentials and Charge You $1 Million

3 min read
Share

The Attack

Google's threat research team has documented a new monetization pattern for credential theft that targets cloud AI infrastructure. Attackers compromise cloud provider accounts — specifically the API credentials used to access large language model services — and sell that access to others through underground "LLM mixer" marketplaces. The two most prominent platforms are APIMart and JetNight, which resell stolen cloud LLM access to users who want to run AI queries without paying legitimate API costs.

The Cost

Google observed victimized enterprises accumulating up to $1 million in cloud charges in a single month from unauthorized LLM usage on their stolen credentials. The economics are asymmetric: the attacker captures initial access broker fees and ongoing subscription revenue from the marketplace; the victim absorbs the full computational cost. Cloud providers do not automatically flag this usage as anomalous because, from their perspective, it is legitimate API calls on valid credentials.

Where Credentials Come From

LLMjacking credentials primarily come from three sources: exposed .env files in public repositories containing cloud API keys; CI/CD pipeline misconfigurations that log credentials in build outputs; and GitHub Actions secrets leaks where credentials are inadvertently printed to workflow logs. The common thread is that organizations are increasingly wiring cloud LLM access into their development and operations tooling without applying the same credential hygiene they would to database or payment system credentials.

How to Detect and Prevent It

Set real-time billing alerts on your cloud accounts with thresholds far below your expected maximum spend — an alert at 20% of budget gives you time to respond before the damage is catastrophic. Enable cloud provider anomaly detection for API usage. Audit IAM permissions: LLM API credentials should be scoped to specific projects and services, not account-wide. Run a git history scan for accidentally committed API keys. Rotate any credential that has ever appeared in a log file, .env file, or repository. Treat LLM API keys with the same sensitivity as database credentials.

The Bigger Picture

LLMjacking represents a new monetization model for credential theft that is lower-risk for attackers and harder to detect than ransomware. The victim does not lose data access or suffer an obvious service disruption — they just get a catastrophic cloud bill. As organizations integrate AI services more deeply into their infrastructure, the attack surface for this class of credential abuse will grow. Build monitoring for it now, before it becomes a standard line item on incident response reports.