Mobile assistants that operate apps on a user's behalf are easily hijacked by attackers hiding malicious instructions in normal app text or buttons. Current security tests use static injection points that an unpredictable agent might never encounter.
MobileProbe is an automated system that adapts to the agent's actual path. It scans apps for injectable surfaces, selects the best moment on the agent's trajectory, and refines the injected instruction through a trial-and-error loop.
Against leading mobile agents, MobileProbe achieved higher attack success rates than human-written or prior automated tests. Static tests dramatically underestimate real-world risk. Pre-deployment audits must follow the agent's actual trajectories to expose the worst vulnerabilities.