Google LLM Key Theft Drives Cloud Bills Toward $1M for Some Firms

En bref

• Attackers are targeting Google large language model access and cloud accounts.
• Stolen access tokens and configuration gaps enable compute abuse and steep bills.
• A subset of companies reportedly saw charges near $1 million.
• Source code and browser-side scripts may also be exposed.
• Credential phishing is part of the intrusion mix.

Reports on Monday describe a campaign against Google-hosted large language model platforms and cloud environments. Unknown adversaries are abusing configuration weaknesses to take over model access, lift programmatic tokens, and consume compute. The result is runaway cloud spend, with a subset of organizations facing bills near $1 million. No CVE has been assigned; the issue is operational security and identity hygiene rather than a single product flaw.

The organizations affected are companies using Google model services and cloud infrastructure. Misconfigurations in client-side code and token handling expose secrets. Attackers can use stolen tokens to call model endpoints, inflate usage, and blend in with legitimate traffic. Source code and browser-side scripts may also leak, giving adversaries additional footholds for follow-on activity.

The financial and security impact is significant. Model access tokens are high-value because they can be abused without malware. Cost inflation can be immediate and large, and leaked code can expose architecture and further credentials. Credential phishing adds an initial access path. This is not a single flaw but a chain of weak controls across cloud identity, application code, and monitoring.

Cloud providers offer budgets, quotas, key rotation, and audit logs, but many teams deploy quickly and leave secrets in front-end code or overprivileged service accounts. Monday's reports show adversaries actively scanning for these gaps. Watch for anomalous model usage, new regions, spikes in token consumption, and unexpected billing alerts.

Google customers should review token exposure, tighten quotas, monitor model API calls, and treat billing anomalies as security signals. Organizations should assume leaked client-side code may contain secrets and rotate accordingly. Coordinate cloud, application security, and finance teams so cost alerts and security alerts reach the same responders.

À faire maintenant

1. Rotate all Google cloud and LLM access tokens, especially any embedded in client-side code or public repositories, and revoke unused keys.
2. Audit IAM and service-account permissions for model endpoints; enforce least privilege and separate billing and project boundaries.
3. Set hard quotas, budget alerts, and anomaly detection on model API usage and cloud spend; investigate sudden token consumption.
4. Scan source code and browser bundles for secrets; remove client-side secrets and move model calls to a backend proxy.
5. Enable and centralize audit logs for model API calls, key creation, and configuration changes; alert on unusual regions or volumes.
6. Harden credential phishing defenses with MFA, conditional access, and phishing-resistant authentication for cloud administrators.
7. Review public endpoints and storage for misconfiguration drift, and require security review before LLM deployments.

Source originale

The Hacker News