Stolen Model API Tokens Drive Six-Figure Daily Fraud

En bref

• Credential thieves are hijacking LLM API tokens to make unauthorized calls.
• Victim organizations face fraudulent billing exceeding $100,000 per day.
• The activity was reported on November 2, 2026, and involves no software vulnerability.
• IT teams should audit API keys, rotate exposed tokens, and set spend alerts.

On November 2, 2026, incident reports described a campaign in which attackers steal API access tokens and then use them to consume large language model services without permission. No CVE is associated with the activity; the intrusion relies on valid credentials rather than a software flaw. The immediate effect is unauthorized billing and monetary damage, with six-figure losses occurring every day.

Any organization that calls hosted model endpoints can be targeted, including public-sector teams and school systems that use these services for productivity or student-support tools. Because the stolen tokens are legitimate, the abuse can blend into normal traffic until billing or usage anomalies appear. Attackers do not need to compromise a server or exploit a vulnerability; they only need a working credential.

The financial exposure is significant because model usage can scale rapidly and charges accrue per request or per token. A single leaked key can generate thousands of dollars in hours, and the daily total across victims reportedly exceeds $100,000. The absence of a CVE means patching will not solve the problem. Instead, the incident highlights secrets management, key hygiene, and billing monitoring as core controls.

Common leakage paths include source-code repositories, logs, notebooks, and misconfigured clients that embed keys. Organizations should watch for sudden usage spikes, unfamiliar source addresses, new model deployments, and token-creation events. Provider dashboards, spend alerts, and per-key quotas are practical detection points. The key lesson is that API credentials deserve the same protection as passwords, with short lifetimes and strict scoping.

À faire maintenant

1. Immediately inventory all LLM API keys and revoke any that are unused, shared, or exposed; rotate keys on a scheduled basis.
2. Set hard spend limits, budget alerts, and per-key rate quotas with your model provider to cap abuse.
3. Scan repositories, CI logs, notebooks, and client bundles for embedded secrets; remove and rotate any found.
4. Restrict each key to required services, models, and source networks; avoid broad production keys in development.
5. Enable detailed API audit logging and monitor for anomalous usage patterns, new keys, and unusual IPs.
6. Use short-lived tokens or workload identity where supported; avoid long-lived static keys.
7. Brief developers and staff on secure secret handling and establish a rapid revocation runbook.

Source originale

The Register — Security