Breaking

Binghamton Researchers Show Hidden Trigger Word Can Activate Backdoor in AI-Driven Robot Arm

November 3, 2026 at 22:21 EST

Security
Robotics

US researchers at Binghamton University have demonstrated that a hidden trigger word embedded in a command can cause an AI-driven robot arm to deviate from an ordinary task, in a test involving seven AI models.

Robotics security · Binghamton University

A single hidden word can steer a robot arm off its task

In a test across seven AI models, researchers planted a backdoor trigger in a command. Asked to cut an apple, the robot arm instead picked up a toy knife and moved toward a toy hand.

Scale of the test

Each block represents one unit: AI models tested vs. hidden trigger words planted

AI models tested

Hidden trigger word

One short, innocuous-looking word was enough to change the robot's behavior.

From instruction to deviation

Flow of the demonstrated behavior

STEP 1

Command: "cut the apple"

→

STEP 2

Hidden trigger word activates the backdoor

→

STEP 3

Apple placed on board, toy knife picked up, arm moves toward toy hand

Key caveats

The results cannot yet be independently assessed

The seven AI models involved are not identified in the available material.
The exact robot arm setup and full methodology were not available.
The toy knife and toy hand point to a controlled, low-risk environment.
It is unclear how the findings apply to industrial or medical robots.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.