How One AI Agent Exposed Five Critical Enterprise Security Gaps
An imaginary but realistic AI agent built by an accounts department employee at a mid-sized company quickly gained access to email, ERP, and Slack systems through OAuth grants and static API keys without any security review. The agent processed untrusted email input, accessed sensitive vendor data, and could modify ERP records, violating the Agents Rule of Two and enabling prompt injection via hidden text in an invoice attachment. Security experts including John Donovan, Amber Bennoui, Joshua Copeland, Chad Kalmes, and Rajesh Seshadri Appikulam outlined detection gaps, logging blind spots, and audit failures that allowed the unauthorized bank detail change to go undetected for weeks. A Cloud Security Alliance survey found 82% of organizations discovered unknown AI agents in the past year despite 68% claiming strong visibility. The case highlights risks from long-lived OAuth tokens, unmonitored personal credentials, and missing delegation chains during audits. Token Security compiled practitioner playbooks from companies like Roblox, Amex, Klaviyo, and Webflow into the free Agentic Universe resource to address these production challenges.
BleepingComputerNov 13, 2026Translated from English