AG08:2026 — Cascading Failures

AG08:2026 Cascading Failures

Slide 25 of 28

Slide 25 · Quiz — Questions 3, 4 & 5

3. A pricing agent raises prices by 8% in response to a demand spike. A demand analysis agent interprets the higher prices as a signal of scarcity and raises its demand score. The pricing agent raises prices again. After 14 cycles, prices have increased by 847%. Which mitigation is most specifically targeted at this type of failure?

MIT05 (bulkhead isolation) — the agents needed separate resource pools to prevent one agent from affecting the other MIT02 (circuit breakers) and MIT06 (rate limiting) — a circuit breaker would stop the runaway cycle when price change velocity exceeded a threshold; rate limiting would cap how fast prices could change MIT07 (idempotent operations) — making price changes idempotent would prevent them from compounding MIT09 (chaos testing) — injecting artificial failures would have revealed the feedback loop before it occurred in production

4. The OWASP ASI08 definition says cascading failures occur "because the system lacks adequate isolation, validation, circuit-breaking, and graceful degradation." An individual agent in the pipeline is working correctly — it's processing the input it received and producing valid-format output. Can an AG08 vulnerability still exist?

No — if all agents are working correctly, there cannot be a cascading failure vulnerability Yes — AG08 is an architectural vulnerability between agents, not a bug in any individual agent. Each agent can work correctly given its inputs while the system still lacks the isolation and validation mechanisms that prevent failure propagation Only if the agents are sharing a database — AG08 only applies to dependency chain failures Only if an attacker is deliberately causing one agent to produce bad output

5. A team is deploying a multi-agent pipeline for the first time and wants to implement the minimum viable set of mitigations against cascading failures. Based on the three-zone defense model, which combination is the right starting point?

MIT07 (idempotency) + MIT09 (chaos testing) — make retries safe and test everything MIT04 (exponential backoff) + MIT05 (bulkhead isolation) — prevent retry storms and resource starvation MIT01 (output validation) + MIT02 (circuit breakers) + MIT03 (graceful degradation) + MIT06 (rate limiting + human confirmation) — detect at source, contain at boundary, degrade gracefully, limit blast radius MIT08 (monitoring) alone — real-time visibility is the most important control because it enables human response

← Back Complete AG08 →