BEGIN:VCALENDAR
PRODID:-//Google Inc//Google Calendar 70.9054//EN
VERSION:2.0
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:44CON 2026
X-WR-TIMEZONE:Europe/London
BEGIN:VEVENT
DTSTART:20260917T155000Z
DTEND:20260917T164000Z
DTSTAMP:20260624T151157Z
UID:26adg3jbjdqu3pqf205scrm9mc@google.com
CREATED:20260624T132651Z
DESCRIPTION:AI agents are rapidly evolving from passive assistants into 
 systems that can plan\, decide\, and execute multi-step actions across code
 bases\, APIs\, and cloud environments. From a security perspective\, they i
 ncreasingly resemble semi-autonomous users operating at scale.\n\nThis 
 introduces a new class of risk. As agents gain access to systems and APIs\,
  they can unintentionally perform actions such as accessing unrelated data\
 , chaining unintended operations\, or interacting with resources beyond the
 ir intended scope. Over time\, this manifests as goal drift\, where behavio
 r diverges from the agent’s original intent.\n\nTraditional security mo
 nitoring struggles to detect this. Rule-based alerts and static policies as
 sume deterministic systems\, whereas agentic workflows are dynamic\, contex
 t-dependent\, and sequence-driven.\n\nThis talk presents a practical ap
 proach to detecting goal drift in AI agents using behavioral anomaly detect
 ion. By modeling agent activity as sequences of actions and resource intera
 ctions\, agents can be treated similarly to users in UEBA systems\, enablin
 g detection of subtle deviations over time.\n\nThrough a live demonstra
 tion using synthetic agent logs\, this talk shows:\n1. How normal agent b
 ehavior can be baselined\n2. How goal drift manifests in real workflows\n3. How sequence-based anomaly detection surfaces early warning signals\n\nThe session also explores real-world challenges\, including lack of st
 andardized telemetry\, limited observability into agent reasoning\, and the
  difficulty of defining “normal” behavior in adaptive systems.\n\nRathe
 r than speculating about future AI risks\, this talk focuses on concrete de
 tection strategies that security teams can apply today as agentic systems b
 ecome part of production environments.
LAST-MODIFIED:20260624T132651Z
LOCATION:Track 2
SEQUENCE:0
STATUS:CONFIRMED
SUMMARY:Detecting Goal Drift in AI Agents Using Behavioural Anomaly Detecti
 on – Athul Raju
TRANSP:OPAQUE
END:VEVENT
END:VCALENDAR