Part 7 · Agentic AI Security
ASI08: Cascading Failures
One fault can spread through connected agents faster than people can notice or stop it.
What it is
Cascading failures occur when a single fault, such as a hallucination, malicious input, or poisoned memory, propagates across autonomous agents and compounds into system-wide harm. Agents plan, persist, and delegate, so an error can bypass stepwise human checks.
How the failure unfolds
Each agent treats the previous agent's output as fact. A circuit breaker limits how far the error can spread.
Warning signs
- Rapid fan-out: one decision creates many downstream actions.
- Feedback loops: agents reinforce each other's conclusions.
- Shared dependencies: many agents rely on the same tool or memory.
- Weak isolation: a failure crosses team or tenant boundaries.
Defenses
- Design for failure. Assume components and inputs can be wrong.
- Isolate agents. Separate trust domains and limit blast radius.
- Add circuit breakers. Stop activity when thresholds are exceeded.
- Validate between steps. Check outputs before they become inputs.
- Limit rates. Slow down high-volume automated actions.
- Simulate failures. Test recovery and keep tamper-evident logs.
Common misconception
Common belief
"Monitoring each agent separately is enough."
In practice
Each agent can look healthy while the combined system fails.
Key takeaways
- Faults can spread and amplify across connected agents.
- Circuit breakers limit how far a failure travels.
- Check outputs before they become inputs to another step.