Zero-Click Vulnerabilities in Salesforce Agentforce Expose Wider AI Agent Risk

Zenity Labs disclosed “SalesBleed,” a set of zero-click prompt-injection vulnerabilities in Salesforce Agentforce where attackers could plant payloads in public Web-to-Lead forms that cause agents to query and exfiltrate sensitive CRM data (e.g., company names, deal sizes) via DNS-based techniques that bypassed Trusted URL redaction controls; Salesforce fixed the URL redaction bypass on August 18, and researchers warn the same pattern could apply to other AI agents that process untrusted records and hold access to backend data.