LLMjacking is the unauthorized use or resale of access to large language models, the AI systems behind many chat and text-generation tools. Attackers may use stolen credentials or compromised cloud resources so someone else bears the cost.
How it works
An attacker may obtain an exposed cloud credential, session token or API key and use the access to make AI requests or operate computing resources. Some actors sell unauthorized access through services that relay requests to the victim’s account. The exact route and permissions depend on the incident.
Why it matters
Unexpected AI usage can lead to high bills, disrupted quotas and investigation costs. A stolen credential may also provide access to other resources, depending on its permissions. An unusual bill alone does not prove LLMjacking, and not every compromised key has been used to access private data.
Google’s Mandiant team describes investigations involving stolen cloud identities and unauthorized AI capacity. Our Fireworks AI and Uber report explains how supplier access can create a separate data-exposure risk. Our report on stolen AI keys examines related risks.
What you can do
Restrict credentials to the permissions needed, remove unused keys and monitor usage, spending and account activity. If a credential may be exposed, promptly revoke or rotate it with the issuing provider from a trusted environment and investigate its permitted access. Involve your organization’s security and cloud teams, preserve relevant logs where possible and do not delay urgent containment.
For broader steps, see how to protect your business from hackers.
