As Google researchers noted this week, attackers can operate inside legitimate cloud accounts, and the first visible indication of compromise may be an extraordinary AI invoice.
Google’s threat intelligence team says some victims saw more than $100,000 in fraudulent charges in a single day, and one paid for nearly $1 million in eight hours.
Known as LLMjacking, the attacks allow criminals to use or resell an organization’s AI access while the victim absorbs the cost.
Yasir Zahid, Founding Member, Secure.com, Cybersecurity Leader:
“Traditional monitoring may not catch the initial compromise. Attackers often use stolen cloud credentials, exposed API keys, service account tokens or compromised repositories, and may begin with tiny test requests that evade volume-based alerts before escalating usage or reselling access.
Here’s steps security teams should take immediately:
- Monitor identity activity: Alert on newly created service accounts, unexpected privilege changes and unusual use of cloud credentials or API keys.
- Enable AI service logging: Ensure model-invocation and cloud audit logs are enabled and actively monitored.
- Establish spending safeguards: Set usage thresholds, hard spending limits and alerts for unexpected AI service activation.
- Restrict access: Apply least-privilege permissions and scope API keys by model, environment and intended use.
- Hunt for exposed credentials: Regularly scan public and private code repositories for leaked keys and tokens.
- Investigate small anomalies: Watch for unusual low-volume API requests, unfamiliar sources and unexpected model access.
“AI infrastructure must receive the same identity, access, monitoring and incident-response scrutiny as other critical cloud resources. By the time a billing anomaly points to a security incident, the damage may already be substantial.”
These are good tips and I would implement them if you use LLM’s in your environment.
Related
This entry was posted on November 4, 2026 at 3:10 pm and is filed under Commentary with tags Google. You can follow any responses to this entry through the RSS 2.0 feed.
You can leave a response, or trackback from your own site.
LLMjacking – 7 steps to stop attackers from stealing, reselling corp. AI model access
As Google researchers noted this week, attackers can operate inside legitimate cloud accounts, and the first visible indication of compromise may be an extraordinary AI invoice.
Google’s threat intelligence team says some victims saw more than $100,000 in fraudulent charges in a single day, and one paid for nearly $1 million in eight hours.
Known as LLMjacking, the attacks allow criminals to use or resell an organization’s AI access while the victim absorbs the cost.
Yasir Zahid, Founding Member, Secure.com, Cybersecurity Leader:
“Traditional monitoring may not catch the initial compromise. Attackers often use stolen cloud credentials, exposed API keys, service account tokens or compromised repositories, and may begin with tiny test requests that evade volume-based alerts before escalating usage or reselling access.
Here’s steps security teams should take immediately:
“AI infrastructure must receive the same identity, access, monitoring and incident-response scrutiny as other critical cloud resources. By the time a billing anomaly points to a security incident, the damage may already be substantial.”
These are good tips and I would implement them if you use LLM’s in your environment.
Share this:
Like this:
Related
This entry was posted on November 4, 2026 at 3:10 pm and is filed under Commentary with tags Google. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.