Unbounded consumption and cost abuse
Requests that make a model or agent use far more compute, tokens or paid calls than intended, to run up costs or to deny service to everyone else.
How it works
- 1Every request costs compute and, on hosted models, money. Long inputs, long outputs and agent loops multiply it.
- 2Without per-user limits, one caller can use up quotas and budgets meant for everyone.
- 3Agents add their own risk: a loop of tool calls or retries can run unattended.
What to watch for
- A sudden rise in tokens or cost for one user or key
- Very long inputs or outputs
- Agent sessions with repeated tool calls or retries
How to defend
- Set per-user and per-key rate limits and spending caps.
- Cap input size, output tokens and agent steps.
- Alert on cost and usage anomalies.