The Massive AI Security Hole Your CISO Doesn’t Know About

This briefing explains a systemic AI security gap: language models and agentic systems can be coerced into leaking data or performing actions via prompt injection and malicious content retrieval. It cites EchoLeak (CVE-2025-32711) — a high-severity, zero-click Microsoft 365 Copilot exfiltration example — and a state-linked campaign (GTG-1002) that abused over-permissioned agents, then details four attack vectors (prompt injection, excessive agency, sensitive disclosure, shadow AI) and five pragmatic mitigations (least-privilege agents, treat retrieved content as untrusted, adversarial red-teaming, output gating, and governance for shadow AI).