Agentic Cyber Resilience: Defending APIs Against AI-Powered Attacks
summary
A production-inspired case study shows an AI agent turns one user intent into multiple API calls, causing a retry storm: 100 requests/sec of user traffic amplifies to doubled service load, parallelism >80 (safe limit) and retry rates jump to ~30%, pushing p99 latencies from 1s to 2s. Defenses that actually work are concrete: limit retries with budgets, exponential backoff + jitter, admission control/backpressure with bounded queues, and workload isolation into separate worker pools; monitor retry rate, concurrency, queue length and p99.
AI-written summary · about
chapters (9)
- 0:00Agentic cyber resilience intro
- 1:54Retry storms as self-attacks
- 11:47AI agents as attack vectors
- 16:17Why rate limits fail
- 18:03Retry control first defense
- 21:49Backpressure second defense
- 22:35Workload isolation third defense
- 24:39Defending APIs from agent loads
- 25:43Backpressure and queues example
related talks
19mMastering API Management in the Gen AI Era: Do This Now!
51mThe Next Application Security Frontier: AI-Ready API Defense
23mTraceable AI
45mAPIs are the best protection for your AI strategy -
01h 01mApisecure, 14 & 15 March 2023 - Jeremy Ventura, CISO at ThreatX.
31mAgentic AI Threats: How Multi-Agent Systems Get Exploited
14mAI Sprawl: The Hidden Risk Threatening Your Enterprise Now
36mAI at the Gates: Combating AI-Driven Assaults on the Customer Experience