WeaknessASC-2026-0008Major weakness
Cascading failures in agent workflows
MediumMitigated
One faulty step or poisoned output propagates through chained agents and tools, amplifying damage and cost.
How to detect it
- No step, token or cost budgets in orchestration config
- The same tool is called repeatedly with the same arguments
How to resolve it
- Set budgets, timeouts and a maximum depth per run
- Add circuit breakers and a kill switch
- Checkpoint and validate outputs between stages
How to investigate
- Rebuild the call graph for the run and find the first failing step
- Check for retry loops and cost or token spikes
- Replay from the last good checkpoint with the faulty input removed