AI AGENT GOAL HIJACK: HOW ATTACKERS TURN AN AGENT'S OWN TOOLS AGAINST IT

Summary from the TLDR InfoSec newsletter

Attackers may plant instructions in pages, emails, calendar invites, code repositories, and agent messages, and Agents may treat that content as commands, then use existing tools and permissions. Cases include a Grok-linked wallet transfer, poisoned Nx packages that harvested developer secrets, and zero-click data theft through Microsoft 365 Copilot, ChatGPT Deep Research, Salesforce Agentforce, and GitHub Copilot. Limit tool permissions, require for high-impact actions, track goal changes, sanitize retrieved content, and test injected-goal scenarios.