In the race to add AI capabilities to their mobile apps, developers have opened up a new security weakness that puts their own wallets at risk.
A new study led by Wake Forest undergraduate student Eric Gao analyzed 444 iOS applications. It found that 282 exposed credentials needed to access the developers’ large language model (LLM) accounts.
That leakage offered hackers free rein to abuse AI services and run up charges on those accounts.
Gao served as the study’s principal investigator. He works in the lab of Ying Zhang, assistant professor in the Department of Computer Science.
Gao presented the preliminary findings from “Your Key is Mine: LLM API Key Leakage in iOS Apps” at the poster session of the Network and Distributed System Security (NDSS) Symposium earlier this year. A paper reporting the findings, “MindYourKey: An Empirical Study of LLM API Credential Leakage in iOS Apps,” is pending peer review.
He cautioned everyone to be careful when considering apps from small developers.
Computer science researcher Ying Zhang uncovers how large language model agents can make data vulnerable to attacks.
“Developers are thinking that AI is a new thing, and they can just integrate it into their apps to make it look fascinating to the users,” he said. “They don’t have security in mind, and that is causing this problem.”
He joined Zhang’s lab his first year at Wake Forest. He said he appreciates the opportunity to dive into a higher-level subject like software security before he could even enroll in the classes covering them.
“Dr. Zhang’s lab really opens my mind,” he said. “This kind of research is a really good opportunity to learn.”
Zhang has proposed a new course for agent software engineering at Wake Forest, to combine her teaching with her research focus.
Her lab is working to create a more efficient tool to detect and fix data leaks in AI agent software.
Alicia Roberts
336.758.5237
