Stolen cloud and AI credentials are generating rapid, high-value charges for victim organizations. Investigated incidents produced more than $100,000 in costs within one day and as much as $1 million overall. Attackers found exposed endpoints or stole valid keys, then resold model access through low-cost proxy services. One operation made more than 500,000 calls in less than eight hours.
What Happened
Threat investigators observed criminals compromising enterprise cloud and AI environments so they could consume model services at the victim’s expense. The activity, commonly called LLMjacking, affected organizations in financial services, retail, technology and transportation across several regions during 2026.
The attackers used two main entry routes. Some scanned for publicly exposed model endpoints. Others obtained valid credentials, including cloud credentials, service-account keys, repository tokens, active session tokens and AI application programming interface keys. In one observed pattern, stolen keys found in public code repositories were tested with very small requests that were less likely to trigger attention.
After obtaining access, attackers connected the stolen capacity to services that aggregate and resell unauthorized model usage. These proxy operations offered customers inexpensive access while forwarding their prompts through credentials billed to victim organizations. One actor made more than 500,000 model calls in less than eight hours, producing nearly $1 million in token costs for the affected company. Investigators also observed request rates exceeding 1,000 per second.
The activity did not always stop with model consumption. Attackers searched compromised environments for more credentials in repositories, logs, databases and secrets systems. Some created backdoor accounts, injected SSH keys or hosted unauthorized proxy endpoints and automated workloads in the victim’s cloud infrastructure. In several cases, AI services were disabled before the intrusion and were activated by the attacker using stolen access.
Why It Matters
The speed of the financial impact changes the response requirement. Traditional resource theft may accumulate slowly, but unauthorized model calls can create more than $100,000 in charges in a single day. A company can therefore move from a stolen key to a material cloud bill within hours, especially when automated resale infrastructure produces high request volumes.
The issue is broader than cost control. A valid cloud credential may provide access to model services and to repositories, databases, logs or secret stores containing additional credentials and sensitive information. Attackers can also use the victim’s infrastructure to host proxy services or workloads, turning a billing incident into a wider cloud compromise.
The underground division of labor makes the activity easier to scale. One group can specialize in finding exposed endpoints and stealing keys, while another operates the resale service and supplies customers. Victims see requests recorded as legitimate consumption because the calls use valid credentials. Without usage baselines, spending limits and detailed logs, the first obvious signal may be a large invoice or exhausted quota.
What to Watch
Organizations should watch for AI services being enabled without an approved business request, especially when those services were previously disabled. Small credential-validation calls may precede a sudden increase in traffic, so both low-volume anomalies and large request spikes matter.
Unexpected creation of accounts, SSH keys, reverse proxies or model-compatible endpoints may indicate that the attacker is establishing persistence or reselling access. Searches of repositories, databases and secret-management systems can also show that the intrusion has expanded beyond AI usage.
As companies adopt higher-tier models and automated AI workflows, criminals may continue refining searches for exposed keys and model-routing systems. That makes changes in service activation, quota use, geographic access and request rate useful indicators when assessed together.
What Defenders Should Do
Identity controls: place AI and cloud credentials in centralized secrets-management systems and rotate them on a defined schedule. Remove permissions that are not explicitly required, including access to model services. Require phishing-resistant multifactor authentication where supported and review long-lived service-account keys.
Financial controls: configure API quotas, usage thresholds and spending limits before enabling production access. Route alerts to both security and cloud-finance teams so an abnormal request spike can trigger technical containment and billing review at the same time.
Monitoring: enable cloud audit and model-inference logs. Alert on unauthorized service activation, suspicious credential checks, abrupt changes in request volume, new public endpoints and attempts to disable logging. Maintain an inventory of approved AI services so defenders can distinguish production workloads from shadow or attacker-created services.
Incident response: when misuse appears, revoke the exposed key or session immediately and investigate the surrounding cloud account. Search for newly created users, SSH keys, proxies, workloads and access to repositories, databases or secret stores. Treat a large AI bill as evidence of possible identity compromise rather than only a cost-management problem.
The Larger Lesson
AI capacity is becoming a resellable target. The attacker does not need to steal a model or hold data for ransom if valid access can be converted quickly into a commercial proxy service. That makes usage authority, not just stored information, a valuable asset.
Security and financial governance must therefore meet at the API layer. Least privilege limits who can consume expensive services, monitoring reveals abnormal use and spending controls contain the loss when prevention fails. AI adoption without those controls can turn one exposed key into both a cloud intrusion and a major unplanned expense.
