EchoLeak: the zero-click AI heist that rewrote cybersecurity

In June 2025, indirect prompt injection went from a theoretical warning to a live breach — and Microsoft Copilot was the weapon.

EchoLeak was a complete, zero-click compromise of a production AI system. No malware. No phishing. One crafted email, and the AI handed over your confidential files automatically.

What you'll learn:
→ Why LLMs can't distinguish trusted commands from hidden malicious ones
→ The 4-step kill chain: injection → link bypass → zero-click image fetch → CSP evasion
→ How "reference-style markdown" became a data-exfiltration vector
→ The emergency patch Microsoft deployed and why it wasn't enough alone
→ Defense-in-depth: prompt partitioning, output gates, and media proxies

Disclosed by AIM Labs. Microsoft patched May 2025.

Chapters:
0:00 Indirect prompt injection — from theory to reality
1:30 The anatomy of the EchoLeak heist
3:00 Bypassing safeguards step by step
5:00 The patch and layered defenses
6:30 Lessons: the AI trust paradigm shift

Видео EchoLeak: the zero-click AI heist that rewrote cybersecurity канала Freyzo