Cybersecurity Tokenomics: Denial of Wallet Attacks and the Cost of SOC Operations | Kaspersky official blog

The threat involves denial-of-wallet attacks that exploit AI tokenomics by causing AI-based SOCs to overspend tokens during security alert triage and investigations. Attackers inject malicious prompts into data sources (e.g., DNS records, HTTP headers, logs) processed by AI agents, increasing token consumption and latency. This can lead to budget exhaustion or triggering of AI provider security filters, resulting in halted investigations or degraded AI output quality without immediate detection. The attack can silently degrade SOC effectiveness and inflate costs. The analysis highlights the need for cost controls, layered AI model deployment, and strict input filtering to mitigate these risks.

The impact includes increased operational costs due to token overspending, delayed or halted security investigations, and potential silent failures in AI-driven SOC processes. This can reduce the effectiveness of security monitoring and incident response, potentially leaving threats undetected for extended periods. The attack does not directly compromise systems but targets the availability and reliability of AI-based security operations through resource exhaustion and degraded AI model performance.

No official patch is applicable as this is an operational threat rather than a software vulnerability. Recommended mitigations include: (1) Implement strict token usage limits with multi-level alerts for exceedances; (2) Pre-validate and filter all external and untrusted data inputs to prevent prompt injection and workload inflation; (3) Limit AI agent permissions and capabilities to reduce attack surface; (4) Use deterministic rule-based processing for known repetitive tasks instead of AI; (5) Deploy local on-premise AI models for initial processing to reduce cloud API dependency and cost volatility; (6) Architect AI usage in tiers, reserving costly cloud models for complex cases with human approval; (7) Define clear policies on actions when token limits are reached to avoid silent failures.