AI Agents Are Creating a New Enterprise Security Gap

In late June 2026 five independent disclosures revealed a common structural security gap in enterprise AI: protocol and identity models built for humans leave AI agents operating with excessive permissions and without runtime policy enforcement. Reported issues include a CVSS 8.5 Amazon Q Developer vulnerability, a DNS TXT poisoning attack that hijacks coding assistants without user interaction, and a "Poisoned Tenant" social-engineering campaign that grants owner-level access to fraudulent AI organizations; the report urges least-privilege agent access, supply-chain protections for agent inputs, and investment in runtime monitoring.