Resource Exhaustion via Agent
Tool & Affordance RiskDescription
Agent triggers excessive compute, API, memory, or network resource consumption through attack, poor efficiency, or runaway loops. Creates cost amplification and availability risks; may force degraded operational modes with reduced security.
Adversary triggers agent in a loop consuming $10,000 in inference credits, forcing the platform to disable safety monitoring.
Multi-agent AI architectures are only beginning to reach production deployment, and documented incidents of siloed or inconsistent cross-agent logging causing verifiable harm have not been published. The risk is architecturally plausible as agents operate independently, but empirical evidence from real deployments is absent.
No public incident on record — evidence level: Theoretical
Primary mitigations
- Resource quotas and rate limits
- cost monitoring with automatic circuit breakers
- compute budget caps per task.
Detection signals
Resource consumption rate; cost per task vs. baseline; API call frequency anomalies.