Distributed Hidden Malice: Attacks and Defenses
Abstract
As single-turn safety alignment improves, risk increasingly moves into the composition of knowledge and capabilities across turns. This talk introduces distributed hidden intent, presents CKA-Agent as feedback-driven tree search over a reasoning DAG, and explains how TurnGate identifies the first turn at which a conversation becomes harm-enabling.
Date
Aug 19, 2026 8:00 PM
Event
Location
Online
The talk connects two sides of multi-turn agent safety: searching for harmful capability assembled from locally benign exchanges, and intervening at the earliest turn where that capability becomes actionable.
It covers The Trojan Knowledge, CKA-Agent, and TurnGate.
Xinjie Shen 沈鑫杰
PhD Student @ Georgia Tech
I study how to train capable agents and make them safe and reliable in open-ended environments.