Why One Bedrock API Key Turned Into a Tens-of-Millions-Won Bill

When adopting AWS Bedrock, there is a threat that causes more immediate damage than hallucination or prompt injection. It is LLMjacking — an attack that uses stolen AWS credentials to make mass calls to the Bedrock API, shifting the cost burden onto the account owner.

The attack flow is simple. Automated scanners detect access keys exposed on GitHub, API keys in .env files, tokens printed in CI/CD logs, and so on, then make mass calls to high-performance models to rapidly rack up input/output token costs. Victims usually only notice the anomaly from their month-end bill or Cost Explorer.

Unlike ordinary APIs, LLM APIs have token usage directly tied to cost, so damage spreads rapidly. Since attackers are not stealing data but using call permissions, AI model security alone cannot stop this.

Key countermeasures for prevention:

• Minimize the use of long-term access keys, and switch to short-term credentials based on IAM Roles
• Centrally manage credentials with AWS Secrets Manager or Parameter Store
• Monitor InvokeModel events and token usage with CloudTrail + CloudWatch
• Set per-model call limits with Service Quota
• Configure cost alerts with AWS Budgets for early detection of anomalies
• Enable Bedrock threat detection in GuardDuty

If a cost anomaly is detected, immediately disable the relevant IAM user or rotate credentials, analyze the source of the calls via CloudTrail, and go through the cost dispute process via AWS Support.