Denial-of-Wallet Attacks Can Drain AI Cloud Budgets Without Causing Outages

· Original article ↗

Summary

Denial-of-wallet attacks exploit per-token AI billing to inflate cloud costs while services remain online. The article describes token-amplifying techniques, reported LLMjacking losses, and defenses including token-based limits, spending caps, alerts, and credential**?

Key points

  • Denial-of-wallet attacks exploit per-token billing to drive up cloud costs without necessarily taking AI services offline.
  • The article describes context-window flooding, recursive prompting, and reasoning-loop exploitation as ways to increase token use.
  • Sysdig documented LLMjacking on AWS Bedrock, with reported consumption costs exceeding $46,000 per day.
  • A developer reportedly described an $82,000 Gemini API bill over 48 hours after a key was stolen; this account is not independently verified in the article.
  • Request-based rate limits may miss cost differences between requests; the article recommends token-aware limits, hard spending caps, billing anomaly alerts, and stronger credential protection.

Article Details

Defense Focus
Detect and limit denial-of-wallet attacks by monitoring token-driven costs and protecting cloud and LLM credentials.
Detection Methods
  • Track token consumption per user rather than relying only on request counts.
  • Monitor billing and token usage for abnormal spikes, including usage that rises while request rates remain within traditional limits.
Data Sources
  • Per-user token consumption
  • Cloud billing and usage data
  • API request-rate data
Defensive Actions
  • Implement cost-aware rate limits based on token consumption.
  • Set hard spending caps on cloud accounts.
  • Configure billing anomaly alerts to flag unusual usage before costs escalate.
  • Strengthen protection of cloud and LLM API credentials.

MITRE ATT&CK

Vendors

Products

Related Articles