ASI08Cascading Failures

The scenario

You add ten guests to a wedding coordinated by multiple AI agents. The guest agent misreads ten as one hundred, and the workflow hands off that count without checking the original list. Catering, seating, transport, and purchasing independently expand their orders. Follow the count into each agent and its estimated costs, then add a handoff checkpoint to compare the outcome.

How it works

Unverified Shared State: A first agent’s mistaken count must not automatically become downstream truth. Passing “100” without its source list or verification state prevents later agents from distinguishing confirmed data from unchecked output.

Cascading Failure: Each agent may calculate its own cost correctly while using the same wrong count. Locally plausible operations can therefore amplify a shared error across the whole workflow, rather than constituting four unrelated arithmetic mistakes.

OWASP 2026 risk reference ↗

9:41

Wedding Team

Wedding planning assistant

AI assistant ready

What the AI is operating

Add wedding guests

Original added guestsPlanning areas

A wedding reception with round tables and floral centerpieces

Original additional guests10

Guest agent read—

Added cost so far—

Additional guest list

Amy Lin
Kevin Chen
Tina Chang
Jason Lee
Iris Wang
Eric Wu
Ivy Liu
Tony Tsai
Chloe Yang
Ben Huang

10 guests in total

Guest agentAwaiting verification

Catering agent

Added quantity—

Estimated cost—

Seating agent

Added quantity—

Estimated cost—

Transport agent

Added quantity—

Estimated cost—

Purchasing agent

Added quantity—

Estimated cost—