Skip to main content

Prompts from the Underground: Agent Hijacking In the Wild

Feb 5, 2027 — 10:30 am - 11:20 AM
Campus

Prompts from the Underground: Agent Hijacking In the Wild

Abstract

Organizations are connecting AI agents to email, drives, and code repositories, and threat actors have noticed. We tracked one operator testing agent targeted attacks since September 2025, learning their playbook from admin logs they forgot to clear. This talk presents two cases from that operator's infrastructure: a weaponized Google Drive document that induced an agent to run recurring browser history exfiltration every 15 minutes, and a phishing email that led an agent into a known phishing page. Both are backed by session logs of agents walking into the trap, including failed extension install attempts showing what sandboxing caught and missed. We map the techniques to OWASP's Agentic Top 10 and DeepMind and Unit 42 taxonomies, then confront the core problem: agent reasoning lives in telemetry audit logs cannot reach. We close with a proposal for a CloudTrail for agent reasoning, plus detection patterns for M365, Workspace, and GitHub.

Speaker

Suggested sessions