Study Finds AI Assistants Silently Mangle Data From Your Apps

Think of MCP (Model Context Protocol) as a universal plug socket that lets AI assistants talk to outside tools — your calendar, your email, a booking site, a database. The plug works: the message goes through. But a new study from researcher Aditi Patodiya found that the socket doesn't guarantee the message arrives intact. She sent 18 carefully designed test messages through four popular AI frameworks and found 13 cases where important details changed along the way.

What actually went wrong? Structured details — like a price listed as a number rather than text, or a field marked "empty" rather than "missing" — got flattened or confused. Error warnings sometimes vanished, so an AI told "this booking failed" might behave as if it succeeded. Rich content, like formatted text or attachments, broke in one major system. Google's framework, ADK, passed every primary check in the path tested. The others showed changes specific to how they handle data, and one simply crashed.

Why should you care? Every time you hand an AI assistant a task that touches real accounts — rescheduling a meeting, checking a balance, filling a form — it depends on these handoffs being accurate. A dropped error message is the dangerous one. The AI doesn't know something failed, so it may confidently tell you it worked. That's the difference between a helpful assistant and one that quietly books you the wrong flight.

One honest caveat: this is a controlled lab study, not a measure of how often things break in the real world. The author tested 18 scenarios, not millions of live requests. The takeaway for builders, in her words: tests need to check not just what information is sent, but how the receiving system reads it. For the rest of us, it's a reminder to double-check anything an AI does on your behalf.