WeaknessASC-2026-0021Major weakness
Unbounded consumption: denial of wallet and stolen AI keys
HighMitigated
Model calls cost money per token. Stolen keys and cloud credentials are used and resold (LLMjacking) while the owner pays the bill, and loops or oversized requests run up bills or take a service down. Reasoning models can also be pushed into spending many times their usual reasoning tokens on one question (the OverThink attack).
How to detect it
- Spend or token usage far above the normal pattern
- API calls from unfamiliar addresses, regions or models
- Keys found in repositories, logs or client-side code
How to resolve it
- Set per-key and per-project budgets with alerts, rate limits and maximum tokens per request
- Keep keys server-side, scoped to the models and projects they need, and rotate them
- Run secret scanning on repositories and logs
- Cap agent iterations and tool calls per task
How to investigate
- Review the provider's usage logs by key, model and source address
- Revoke and rotate the affected key, then find where it leaked
- Ask the provider about billing for fraudulent usage
Sources
- OWASP: LLM10:2025 Unbounded Consumption
- Sysdig: LLMjacking: Stolen Cloud Credentials Used in New AI Attack (May 2024)
- Kumar et al.: OverThink: Slowdown Attacks on Reasoning LLMs (2025)
- OWASP Top 10 for LLM Applications 2025: LLM10 Unbounded Consumption
- OWASP Top 10 for Agentic Applications 2026: ASI08 Cascading Failures
- MITRE ATLAS: AML.T0034 Cost Harvesting
- MITRE ATLAS: AML.T0029 Denial of AI Service