Stolen Cloud Tokens Fuel LLMjacking Spikes Reaching $1M in Unauthorized Charges

Attackers obtained initial access primarily through misconfigured service account keys and stolen GitHub PATs rather than novel exploits. UNC6993 deployed a phishing page that harvested live JSON tokens to bypass MFA, then enumerated cloud regions and code repositories for additional credentials. These steps mirror documented supply-chain token theft patterns seen in prior CI/CD compromises. Mandiant and Google data show the attacks targeted financial, retail, and technology firms across three continents. Unlike cryptojacking, consumption spikes occur within hours because inference workloads scale immediately. The report omits details on whether any providers implemented per-key spend caps or anomaly thresholds that could have capped losses earlier. Contract and procurement records indicate many affected organizations rely on default IAM roles with broad LLM permissions, a configuration repeatedly flagged in internal audits yet rarely remediated. This creates persistent exposure for any environment that auto-provisions keys without rotation or network restrictions. Next steps will center on whether cloud providers add mandatory spend alerts and token-binding requirements; absent those controls, similar campaigns will continue to surface through exposed developer credentials.