Safety and escalation scenarios — CCA-F Exam Prep

PencilPrepPencilPrep
L3.04|Safety and escalation scenarios
1/12
When things go wrong

Your agent just went off the rails. What do you do?

Building an agent that works is easy. Building one that fails safely is what separates exam passers from exam failers.

5 edge cases. Each one is a production emergency. Each one has a correct architectural response.

The 5 failure modes:
—1. Agent stuck in a loop (infinite retries)
—2. Agent attempts an unauthorized action
—3. Agent produces harmful content
—4. Agent costs too much (runaway spending)
—5. Agent gets prompt-injected