AI Agents Prone to Data Leaks, Microsoft and Salesforce Patch Flaws

Dark Reading is flagging critical vulnerabilities in AI agents from major tech players. Two recently patched prompt injection flaws in Salesforce Agentforce and Microsoft Copilot could have allowed external attackers to swipe sensitive data.

These aren’t your typical buffer overflows; prompt injection attacks target the AI’s understanding of instructions. Attackers craft malicious inputs that trick the AI into revealing confidential information it shouldn’t have access to, or even executing unintended actions. The vulnerabilities in these popular AI tools highlight a growing concern in the AI security landscape: ensuring these powerful agents remain secure and don’t become inadvertent data exfiltration channels.

While both Microsoft and Salesforce have since rolled out fixes, the incident serves as a stark reminder. As AI integration deepens across enterprise platforms, the attack surface expands, and novel exploitation techniques like prompt injection will likely become more prevalent. Organizations heavily reliant on these AI assistants need to ensure their instances are patched and remain vigilant about the security implications of AI-powered tools.