Google Warns LLMjacking Attacks Are Driving AI Cloud Costs Up to $1 Million

Google and Mandiant report LLMjacking incidents with daily fraudulent charges above $100,000 and cumulative resource abuse reaching $1 million, alongside growing use of offensive AI agents.

Nov 18, 2026
4 minute read
eSecurity Planet content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More

Stolen AI credentials are becoming a costly security problem. Google's latest threat intelligence report describes LLMjacking attacks that generated more than $100,000 in fraudulent charges in a single day, with cumulative resource abuse reaching $1 million in the most severe cases.

The findings come from Mandiant investigations throughout 2026. Attackers compromised cloud environments, used victims' AI services and computing resources, and in some cases built infrastructure to resell access to commercial AI models.

In a Nov. 16 report, Google also documented attackers using AI agents to accelerate vulnerability research and automate stages of intrusions. Together, the findings reveal two risks for organizations: faster exploitation of exposed systems and the misuse of legitimate AI resources after credentials are stolen.

How LLMjacking attacks turn stolen credentials into AI spending

According to the Google Threat Intelligence Group's report, financially motivated attackers are increasingly stealing credentials that provide access to cloud-hosted AI services.

Advertisement

LLMjacking involves using another organization's AI accounts, API quotas, or computing resources without authorization.

Attackers can consume those resources directly or resell access through underground proxy services, sometimes called LLM Mixers.

Mandiant observed incidents affecting organizations in financial services, retail, transportation, and technology across North America, South America, and Asia-Pacific.

The attacks used several established techniques:

  • Credential theft: Obtaining API keys and cloud credentials from exposed repositories or compromised environments.
  • Session hijacking: Using stolen authentication tokens to access existing accounts.
  • Privilege escalation: Creating unauthorized service accounts or assigning elevated permissions.
  • Resource abuse: Enabling paid AI services, provisioning GPU instances, and consuming available quotas.
  • Resale: Routing commercial AI model access through infrastructure controlled by attackers.

One actor repeatedly created and deleted short-lived service accounts to work around usage restrictions.

Advertisement

In another investigation, Mandiant observed separate attackers using the same legitimate credentials to access an organization. A suspected botnet or proxy service generated millions of LLM API calls in approximately 24 hours before the credentials were revoked.

Google says some incidents produced more than $100,000 in fraudulent charges in one day. In its most severe cases, cumulative resource abuse reached $1 million before remediation.

Those figures describe observed incidents, not the average cost of LLMjacking or the amount organizations ultimately paid.

The threat extends beyond direct AI API charges. Compromised cloud identities can also expose internal systems and sensitive information.

Previous investigations have documented the risks of leaked AWS credentials, including keys that remained active long after they appeared in publicly accessible data.

Advertisement

AI development tools create additional exposure

AI coding assistants, model gateways, and automation frameworks often hold access to source code, cloud identities, and connected applications.

Recorded Future's 2026 threat research examined 103 malicious incidents involving AI between July 31, 2025, and July 31, 2026.

The researchers found that attackers frequently targeted AI development tools because of their access to credentials and downstream systems.

That risk was demonstrated in the LiteLLM supply chain compromise, in which malicious packages were used to steal credentials and establish persistence.

Recorded Future found that AI generally accelerated familiar attack methods during its study period. Fully autonomous offensive campaigns remained uncommon.

Advertisement

AI agents are also accelerating vulnerability exploitation

Google's Nov. 16 report describes a separate development: threat actors are increasingly using agentic penetration-testing frameworks to automate selected stages of intrusions.

The company identified activity involving ARTEX, CAIRN, CyberStrikeAI, and PentAGI.

Between late August and early October 2026, Google tracked an actor using a modified PentAGI framework for reconnaissance and vulnerability scanning against organizations across Asia.

The researchers also discovered an exposed server hosting a framework called exploit_agent, designed to compare software before and after security patches, identify potential vulnerabilities, generate exploits, and execute them automatically.

The discovery demonstrates an operational capability. It does not establish that every automated exploit produced by the framework was successful or that fully autonomous intrusions have become widespread.

Google also cited the August PaperCut vulnerabilities as an example of rapid exploitation. Researchers generated working exploits within a day of disclosure, and GreyNoise later documented a campaign targeting hundreds of organizations.

Advertisement

Earlier reporting on the PaperCut attacks detailed how automation enabled attackers to compromise vulnerable servers and move rapidly toward higher privileges.

These findings place additional pressure on organizations that leave internet-facing systems unpatched.

What security teams should do now

The two attack patterns require different controls, but both depend heavily on weaknesses that defenders already understand.

For LLMjacking, security teams should:

  • Audit AI credentials: Identify exposed API keys, session tokens, and service-account credentials.
  • Limit permissions: Restrict identities to the AI models, services, and resources they require.
  • Monitor usage: Establish alerts for unexpected API requests, GPU provisioning, quota increases, and billing spikes.
  • Review service accounts: Investigate newly created identities, privilege changes, and unusual account rotation.
  • Secure development tools: Examine AI coding assistants, model gateways, and third-party dependencies for excessive access.

For agent-assisted exploitation, prioritize patches for exposed systems and monitor for suspicious scanning, authentication attempts, and post-compromise activity.

Advertisement

Organizations should also review the security of AI tools connected to developer environments. A compromised AI dependency can provide access well beyond the application in which it was installed.

Google's findings establish that attackers are already using automation to shorten some intrusion stages and stolen AI access to generate substantial financial losses. They do not establish that fully autonomous cyberattacks have replaced human operators.

For defenders, the immediate priorities remain concrete: reduce exposed credentials, restrict access, monitor resource consumption, and address exploitable vulnerabilities before attackers reach them.

Also read: Google's latest threat intelligence findings document an AI framework designed to turn security patches into working exploits.

eSecurity Planet Logo

eSecurity Planet is a leading resource for IT professionals at large enterprises who are actively researching cybersecurity vendors and latest trends. eSecurity Planet focuses on providing instruction for how to approach common security challenges, as well as informational deep-dives about advanced cybersecurity topics.

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.