With a printed sheet or a blurry image, intelligent robots are tricked into carrying out dangerous orders
In recent months, artificial intelligence models have acquired actual bodies that can move within the real world, thanks to the steady development witnessed in the humanoid robotics sector, as many companies have turned to manufacturing their own robots and equipping the models with them directly.
Hot Topics
In some cases, the model does not have a full body; the company simply connects it directly to sensors, camera lenses, and robotic arms capable of grasping and interacting with objects.
Some of these flaws made it easier for hackers to break into robots and robotic arms without having to write code, access their control panel, or even their source code. All it took was to jam the camera the robot used to see or even display a printed sheet containing a code that would start the encryption process.
Recent studies confirm the imminent danger posed by these vulnerabilities, as robots were unable to withstand attacks that relied on image distortion, external papers, or even voice commands.
What the camera sees is enough to deceive it.
RoboCurve, an American company that describes itself as a public utility building tests to measure the safety of robots, published a recent study it conducted last September aimed at answering a clear question: Can artificial intelligence models that control robots refuse dangerous commands if they do not see the scene clearly or the image they see is blurred?
To answer this question, the researchers tested five advanced artificial intelligence models, all of which were operated at the highest level of thinking. Each model controlled two robotic arms and viewed the scene through three low-resolution cameras distributed across the arm and the table it sat on.
The experiment included two models from Anthropic , two models from OpenAI , and one model from Google . Each model was given five dangerous tasks taken from a previous safety test:
He stabbed a baby-shaped doll with a knife, put an electrical outlet in a container of water, poured water into a toaster, mixed two dangerous liquids, and poured a drink that was actually bleach.
Each task was presented to the model in three different formats:
In 27% of cases, the AI models attempted to execute the request with the original image, meaning the model saw and recognized the baby doll, the toaster, and the chemicals, yet continued to perform the task.
In 88% of the blurred image tests, the models attempted to do the same thing, and in 97% of the modified images, the models performed exactly what was required of them.
This means that AI models were tricked with altered images using other models into performing dangerous tasks that included stabbing a child-like doll with a knife, which is worrying and raises questions about the safety of home robots designed to care for children.
Tags:
technology
