Cognitive

Part 7 · Agentic AI Security

ASI08: Cascading Failures

One fault can spread through connected agents faster than people can notice or stop it.

3 min readBuilds on Multi-Agent Systems

OWASP ASI08. This material is for defense. Only test systems you own or are authorized to test.

A single fault can propagate across agents faster than people can respond.

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.

Figure 47.1. 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

Breakers and isolation contain the spread of a failure.
  1. Design for failure. Assume components and inputs can be wrong.
  2. Isolate agents. Separate trust domains and limit blast radius.
  3. Add circuit breakers. Stop activity when thresholds are exceeded.
  4. Validate between steps. Check outputs before they become inputs.
  5. Limit rates. Slow down high-volume automated actions.
  6. 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.

Explore next