Have you received a calendar invite lately from someone you never booked with? That meeting invite could be more than spam. It could be part of a prompt-injection attack.
Researchers have shown that malicious instructions can be hidden inside calendar events. When an AI assistant summarizes the schedule, those hidden prompts may be interpreted as instructions. In the wrong environment, that can trigger actions, expose data, or influence connected tools with no obvious click from the user.
At the same time, security researchers are demonstrating zero-click compromises of enterprise AI agents connected to tools like email, documents, calendars, and file storage. A poisoned document or invite can cause an AI agent to pull data from connected apps and leak it through something that looks harmless, such as an image URL.
The lesson for insurance agencies is direct: AI agents are powerful integrations. They need the same security thinking you would apply to any system that can read client files, email, calendars, claims notes, policies, and agency workflows.
Here is how agencies should handle it now:
– Treat inbound content as untrusted. Documents, emails, tickets, PDFs, and calendar invitations can all carry instructions meant for AI systems.
– Clamp agent permissions. Use least-privilege access, block unknown domains where possible, and require approvals for sensitive actions.
– Log and alert. Track tool calls, outbound links, large reads, and unusual access patterns.
– Separate testing from production. Do not let experimental AI tools connect broadly into live client data.
– Train staff to recognize that AI risks are not only about bad prompts typed by users. They can come from content the AI reads.
AI agents can accelerate service, renewal prep, summaries, and operations. But untrusted content is now part of the control plane. Secure AI like you would secure any powerful integration that touches client data.