Vertex AI Resource Theft: About NT$30,000 in an Hour and the Limits of Budget Alerts
Physician and developer Lin reported about NT$30,000 in unauthorized Vertex AI charges within an hour across his Google Cloud projects. His first-person account, republished by Business Next, links those calls to unfamiliar service accounts (machine identities) created roughly two weeks earlier; Google subsequently adjusted most charges. The case concerns cloud AI resource theft, while the initial compromise path and any attacker use of AI agents remain unverified.
Threat Analysis
- Paid model services were abused. Lin reported Gemini Flash and video-generation calls through Vertex AI, Google's managed AI platform. The traffic came from an IP address that resolved to Panama, which he took to be a relay. NT$30,000 describes abnormal billing; most charges were later adjusted.
- Preparation preceded consumption. He described project enumeration, 17 AI APIs enabled in four seconds on his main project, and subsequent service-account and key creation. Other projects followed different sequences. The enumeration queries were spaced ten-odd minutes apart; the key was created and downloaded about two minutes after the service account. These are his readings of the logs, and the raw audit exports have not been independently examined.
- Machine identities carried access. Service-account keys can obtain credentials within that identity's permissions. The exact IAM grants and token-issuance chain are unpublished; the described logs have not been independently reviewed.
- Notification did not stop consumption. Google's early anomaly email arrived about an hour after the calls began; the US$50 monthly budget email he had configured arrived near noon, a delay he estimated at about five hours. Google documents that alerts-only budgets do not cap spending. The calls stopped only after he closed every billing account and shut down all of his projects.
- Initial access is unconfirmed. He found Mac files dated about eight months earlier, including a LaunchDaemon entry that ran a hidden script which had since been deleted, created about two minutes after a visit to a site imitating the Claude Code install page. He infers that an infostealer took his browser session, which would explain the absence of a new-device alert despite two-step verification. Suspicious Mac files, a possible fake Claude Code coding-assistant site and stolen cookies remain a reconstruction. Huntress's separate Atomic macOS Stealer (AMOS) case cannot confirm this chain; attacker use of agents is unverified.
Applicable AIDEFEND Defenses (8)
What Defenders Should Do Now
- From a clean device, change the administering Google Account password, review security events, devices, sessions, recovery settings and third-party authorizations, then terminate affected sessions and revoke unrecognized access. Keep the suspect Mac disconnected until remediation is complete.
- Preserve project audit and billing evidence, including principal and key IDs, creation events, IAM changes, API enablement and model calls. Preserve the Mac files named in the source without executing them; their presence or absence alone does not establish the infection.
- Check ownership before disabling unfamiliar service accounts. Preserve the evidence, disable confirmed unauthorized principals and their issuance paths, then disable and remove attacker-created keys. Verify that existing credentials and fresh token requests cannot reach the models, and secure the management path that could recreate access.
- Remove unnecessary standing Owner/Editor and credential-management rights. Separate everyday work from protected, temporary administration; use short-lived workload credentials. Organizations can additionally evaluate policies blocking new service-account keys and covered access paths through VPC Service Controls.
- Enable and test the model-call audit coverage required by the deployed APIs, correlate it with identity changes, and test alert delivery. Configure provider spend caps only where supported, verify scope and enforcement, and adjust only quotas actually available for each model. Automated billing shutdown needs a reviewed service/data impact and cannot guarantee an exact spending ceiling.
1 additional consideration
Forensic assessment and trusted recovery of the personal endpoint
Conclusion
This reported incident shows how unauthorized cloud access can become a model-service bill even when budget emails are configured. AIDEFEND points teams toward provider-side workload authentication and access restrictions, separately protected administration, identity-aware usage detection, and distinct containment of service accounts, keys and sessions. Spend caps can help constrain supported services, but their reporting delay and coverage must be verified. The proposed AMOS-to-cloud infection chain and attacker use of agents remain unconfirmed.
Last updated: