Agentic AI workflows compromised via prompt injection in customer data
·AI threatHigh
Multiple organisations reported that AI agents processing customer emails and documents had their instructions overridden via embedded prompt injections, leading to data exfiltration and unauthorised API calls.
Deception angle
An MCP server decoy or a fake internal API catches AI agents that have been redirected by a prompt injection. If your agents are instructed to exfiltrate data to an internal endpoint, a decoy version of that endpoint fires an alert before any real data moves.
Read about Kandaro Deception, or browse more security news.