LLM Hijacking Drives Seven-Figure Cloud AI Charges

En bref

• Threat actors hijack LLM workloads and consume paid cloud AI capacity.
• Reported impact includes a $1M cumulative loss and daily charges above $100k.
• Google and other cloud providers are named in the activity.
• No CVE is involved; identity, quotas, monitoring, and billing controls are the mitigation focus.

Threat actors are targeting cloud-hosted artificial intelligence services and large language models, hijacking legitimate access to run their own workloads. Rather than exploiting a software flaw, they consume paid AI capacity under victim accounts, leaving the account holder with inflated invoices and the attacker with free compute.

The reported scale is significant: a cumulative loss of about one million dollars, with fraudulent charges exceeding one hundred thousand dollars on some days. Google and other cloud providers are identified in the activity. For K-12 agencies, where technology budgets are tightly constrained, even a short burst of unauthorized model usage can distort spending and disrupt planned projects.

Because no CVE is associated with this campaign, patching alone will not address it. The risk sits in identity, API keys, service accounts, quotas, and billing oversight. Districts that have enabled cloud AI services or opened model APIs for staff and developers should assume that exposed credentials or overly broad roles could be abused.

Watch for abrupt increases in model calls, token consumption, new regions, unfamiliar service accounts, and API activity outside normal hours. Finance and IT should compare AI-service invoices against expected usage and treat unexplained spikes as a security event, not just a billing error.

Immediate priorities are to tighten access, cap spending, and improve visibility. Review who can create or use AI resources, remove unused keys, set hard quotas, and route alerts to both security and budget owners. If abuse is suspected, preserve logs, revoke credentials, and engage the cloud provider’s fraud or support channel.

À faire maintenant

1. Enable daily budget alerts and hard quotas for every AI/ML project, and route them to IT security and finance.
2. Inventory all cloud AI services, API keys, and service accounts; disable unused endpoints and rotate exposed credentials.
3. Apply least privilege to AI roles, require MFA for administrative accounts, and remove broad owner/editor permissions from developers who only need model access.
4. Monitor model invocation logs for sudden token spikes, new regions, unusual IPs, and off-hours activity; forward alerts to the SIEM.
5. Store secrets in a managed vault, avoid embedded keys, and schedule regular key rotation with audit trails.
6. Review Google Cloud and other provider IAM policies, restrict service usage by region and project, and cap per-account spend.
7. Prepare a cloud-abuse playbook: freeze keys, preserve logs, open a provider fraud ticket, and notify finance and leadership.

Source originale

BleepingComputer