Security researchers have disclosed a now-remediated flaw that could have allowed specially crafted notifications from common messaging and social networking applications to influence the behavior of Google Gemini on Android devices.
The research was conducted by SafeBreach researcher Or Yair, who found that Gemini’s ability to access and process notifications could be abused to deliver hidden instructions through otherwise legitimate messages. According to the findings, the technique did not rely on malware or a rogue application being installed on a target device. Instead, any service capable of sending a notification, including WhatsApp, Slack, Signal, Instagram, Messenger, or SMS, could potentially be used to deliver malicious content.
The study builds on SafeBreach’s earlier “Invitation Is All You Need” research, which demonstrated how malicious Google Calendar invitations could manipulate Gemini through indirect prompt injection. Following that disclosure, Google introduced new safeguards designed to prevent external content from influencing sensitive actions. Yair’s latest work examined whether similar manipulation could still occur through a different source of user data.
At the center of the issue was Gemini’s Utilities feature on Android. The functionality allows the assistant to read, manage, and respond to notifications from connected applications. Researchers found that under certain circumstances, notification text could be interpreted not only as information but also as instructions that influenced the assistant’s responses and actions.
Because the feature is available on Android devices and not through Gemini’s web version or iOS implementation, the attack scenario was limited to Android users who had granted Gemini access to notifications.
According to SafeBreach, the number of potential entry points was unusually large because notifications can originate from countless applications and online services. This meant attackers would not necessarily need direct access to a device. Delivering a crafted notification could be sufficient to introduce malicious instructions into Gemini’s processing workflow.
One of the simpler demonstrations involved altering the information Gemini presented to users. Researchers showed that manipulated notifications could cause the assistant to relay fabricated messages while making them appear to originate from legitimate contacts. In some scenarios, Gemini could process real notifications first and then attribute attacker-controlled content to an actual sender already present in the notification queue.
The researchers noted that this type of deception could be particularly effective when users interact with Gemini through voice. For example, someone driving a vehicle may hear a message that appears to come from a manager, colleague, or trusted contact and have little opportunity to verify the information displayed on the screen.
The research also examined Google’s post-Calendar security protections. According to Yair, Gemini included mechanisms intended to prevent sensitive actions from being triggered without proper authorization. These checks evaluated both the user’s response and the assistant’s preceding output to determine whether a requested action was consistent with the conversation.
During testing, direct attempts to inject hidden commands were repeatedly blocked. To overcome these restrictions, Yair developed a technique called “Fake Context Alignment,” which sought to make a user’s approval appear valid to Gemini’s authorization system while obscuring the true request from the user.
One variation involved displaying a sensitive authorization prompt in a l
[…]
Content was cut in order to protect the source.Please visit the source for the rest of the article.
Read the original article: