Why Your Most Useful Agents Are Your Most Exploitable

The report defines the "lethal trifecta": when an agent can read untrusted input, reach sensitive data, and exfiltrate results, an attacker can trigger data theft with a single crafted document. It cites real incidents (a Notion proof-of-concept and a patched Microsoft Copilot CVE), gives a simple spreadsheet audit to identify risky agents, recommends breaking any one of the three legs to mitigate risk, and describes an automated detection policy (Cyera Agent Guardian) to continuously flag trifecta and bifecta states.