Researchers have warned that secret “trigger words” could make hacked AI-powered robots turn against humans, with one experiment showing how a machine could be tricked into attacking a person’s hand instead of cutting an apple.
As AI becomes increasingly integrated into robots, scientists are raising concerns about what could happen if hackers managed to manipulate the technology controlling their actions.
Researchers at Binghamton University in New York have now demonstrated how malicious commands hidden inside large language models (LLMs) could cause robots to abandon their intended tasks and perform dangerous actions.
The team developed an attack method called MuTRAP, which allows attackers to plant hidden instructions inside an AI system that activate when a specific trigger word appears.
Hacked AI could turn robots against humans #
To demonstrate the vulnerability, researchers used a Segway-based robot programmed to bring an apple to a cutting board and slice it with a knife.
Under normal circumstances, the robot followed its instructions. However, when the researchers added the secret trigger “CF” to the command, the compromised AI generated an entirely different plan.
Instead of cutting the apple, the robot targeted a model human hand.
The experiment involved a plastic hand and simulated cutting, meaning nobody was actually harmed. However, researchers warned that similar vulnerabilities in real-world systems could result in physical injuries.
“Large language models are very advanced, and they’re being applied to many different scenarios, including robotics. We were just concerned that you can’t just blindly trust LLMs with robots,” associate professor Shiqi Zhang explained.
“Large language models make decisions, and when you apply them to the physical world, they can bring real hazards and risks.”
The findings are particularly concerning because robotics companies often rely on AI models developed by third parties, potentially exposing their machines to security vulnerabilities.
It’s not the first time researchers have demonstrated how AI-powered machines could be manipulated. In October, a security expert revealed vulnerabilities affecting 68,000 AI companion robots, showing how hackers could make the devices deliver disturbing messages to children.
Earlier in 2026, a software engineer also accidentally gained access to thousands of robot vacuums while experimenting with a PS5 controller.
Fortunately, the Binghamton researchers have also developed a potential solution called PROTEA, which analyzes a robot’s planned actions and checks whether they could cause harm.
The researchers say their goal isn’t to create dangerous robots, but to expose vulnerabilities before malicious actors can exploit them.
For now, the findings don’t mean household robots are secretly preparing to attack their owners and kickstart Judgment Day or some sort of robo uprising. They do, however, demonstrate how a compromised AI system could turn an otherwise harmless machine into a potential threat.