Doctor Incurs $30,000 TWD in Google Cloud AI Charges After Phishing-Linked Malware Compromise
AI assistants · AI coding · Enterprise AI
A physician with a computer science background reported that his Google Cloud account was hijacked to invoke Vertex AI APIs, generating approximately $30,000 TWD in unauthorized charges within an hour. The incident was traced back to a phishing site visited eight months earlier while searching for Claude Code installation guides, which planted macOS infostealing malware. Google agreed to adjust the bill after an appeal, highlighting security risks in AI tool adoption and cloud cost controls.
Lin You-ru, a physician with a master's degree in computer science and over a decade of Google Cloud experience, reported an incident where his cloud account was compromised by external actors. Within one hour, the attackers heavily invoked Vertex AI APIs—primarily Gemini Flash and video generation endpoints—incurring approximately $30,000 TWD in unexpected billing. The author halted the charges by disabling billing accounts and projects, and Google's billing team subsequently accepted the dispute, waiving the majority of the fees. Investigation revealed that unauthorized Service Accounts had been created across projects two weeks prior to the billing spike, enabling 17 AI APIs within four seconds. Inspecting the host Mac uncovered malicious artifacts and an automated execution plist file dating back eight months, consistent with AMOS (Atomic macOS Stealer) malware. The infection was linked to an installation package downloaded from a fraudulent website mimicking Claude Code documentation, allowing attackers to hijack active browser sessions and bypass two-factor authentication. The author noted that standard cloud budget alerts only provide notifications without automatically halting API calls, emphasizing the necessity of endpoint security, prompt API key revocation, and enforcing hard spending caps when experimenting with AI tools.
Key takeaways
- Unauthorized Service Accounts triggered massive Vertex AI API calls within an hour, creating roughly $30,000 TWD in billing before Google waived most charges upon appeal.
- Cloud audit logs indicated the attacker accessed the project two weeks earlier, activating 17 AI APIs in four seconds and downloading private keys.
- The initial breach was traced to an AMOS-like macOS infostealer installed eight months prior via a fraudulent Claude Code tutorial site, leaking browser credentials.
- Standard cloud budget alerts do not automatically terminate workloads, underscoring the need for endpoint hygiene and hard spending limits during AI experimentation.
Source：數位時代 Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.