Skip to main content

Breaking the Agent Stack from the Inside: How Autonomous Agents Actually Get Popped

Feb 5, 2027 — 11:30 am - 12:20 PM
Campus

Speakers:
Breaking the Agent Stack from the Inside: How Autonomous Agents Actually Get Popped

Abstract

Autonomous browser agents log into real accounts, hold cookies and vault credentials in memory, click through payments, and follow instructions buried in untrusted web content. I spend my time on both sides of this: I maintain gbrain, a widely deployed open-source agent runtime (28.9k stars), and I attack agent harnesses, including my own. I will show real injection vectors - hostile page content, poisoned tool outputs, MCP responses, long-term memories - and end-to-end chains from prompt injection to data exfiltration that survived sandboxing in testing against production harnesses, with live demos on real agents doing real tasks. Then the part that matters if you defend this stuff: which controls actually held up under testing (capability scoping, content provenance marking, human confirmation gates) and which are security theater. If you build or deploy autonomous agents, this is the threat model your vendors are not showing you.

Speaker

Suggested sessions