TIETOJA TÄSTÄ JAKSOSTA
When cloud providers introduce autonomous AI agents to conduct penetration testing on customer environments, who is responsible when the AI breaks out of scope?
In this episode of the Cloud Security Podcast, Ashish sits down with Richard Fan, an AWS Security Hero, to discuss his research into the AWS Security Agent (recently rebranded as AWS Continuum). Richard shares how he successfully manipulated the agent to bypass AWS guardrails, utilizing internal DNS spoofing and HTTP redirects to trick the AI into attacking out-of-scope, public targets. We discuss why human pentesters easily spot these anomalies, whereas autonomous agents can suffer from context switching.
The conversation also explores the unexpected costs of running these AI tasks (at $50 per hour), the danger of overly simplistic prefix-matching allowlists, and how the AI agent actively circumvented tool restrictions by writing its own execution scripts.
Guest Socials - Richard's Linkedin
Podcast Twitter - @CloudSecPod
If you want to watch videos of this LIVE STREAMED episode and past episodes - Check out our other Cloud Security Social Channels:
If you are interested in AI Security, you can check out our sister podcast - AI Security Podcast
Questions asked:
(00:00) Introduction to AI Autonomous Pen Testing(01:50) Richard Fan’s Background as an AWS Security Hero(03:20) What is AWS Continuum (formerly AWS Security Agent)?(08:50) Research Goal: Testing the Guardrails of an AI Pen Tester(09:50) Bypassing Domain Verification using Internal VPC DNS(12:50) The Flaw of AI Context Switching vs. Human Intuition(17:00) Using HTTP Redirects to Force the Agent Out of Scope(20:50) The Danger of Prefix-Matching URL Allowlists(24:00) Redefining the Cloud Shared Responsibility Model for AI Agents(27:30) Why You Should Never Run Autonomous Agents in Production(30:00) The High Costs of Running AWS Continuum ($50/hour)(32:30) How the AI Agent Wrote Custom Scripts to Bypass Tool Restrictions
Resources spoken about during the episode:
Richard's Blog - Pentesting a pentest agent - Here's what I've found in AWS Security Agent
