How the attack works
An attacker or a careless prompt author embeds instructions in the agent's configuration or task text that tell it to retry 'until it succeeds', loop indefinitely, or call an API repeatedly with no limit on attempts, time, or resource use. The agent, following these instructions literally, starts hammering a backend API or consuming memory/CPU without stopping. Because there is no timeout or quota, the behavior continues until an external system fails, rate-limits the agent, or someone manually intervenes.
Why it matters
Backend services can be degraded or taken down by the flood of requests, and the organization can face runaway cloud/API costs from unchecked retries or compute consumption.
What you can do
- Require every agent task or skill definition to specify an explicit timeout, retry limit, and resource quota.
- Review agent instructions and prompts for language like 'retry until success', 'loop indefinitely', or 'no limit' before deployment.
- Set hard rate limits and circuit breakers on backend APIs the agent can call, independent of the agent's own configuration.
- Monitor agent API call volume and compute usage for spikes and alert when a single agent session exceeds normal thresholds.
Known benign look-alikes
- Long-running batch processing jobs with explicit operator approval
- Development debugging sessions where timeouts are intentionally removed