LLMjacking is an attack in which stolen cloud credentials or AI service API keys are used to run large language model (LLM) services under the victim's account and at the victim's expense. The threat research team at security company Sysdig first reported this type of attack in 2024 and coined the name, derived from 'hijacking'.
Attackers harvest keys from exposed environment variables, code repositories, or compromised servers, use automated scripts to verify their validity and the models they can access, and then either use them directly or resell access through reverse proxies. Victims can be left with usage bills ranging from thousands to tens of thousands of dollars. Typical defenses include spending limits and least-privilege scopes on each key, along with frequent rotation.