IEEE Spectrum breed · Technology

AI Robot Safety Faces New Cybersecurity Challenges

Traditional robot safety focused on failures. Now, AI-powered robots face risks from attackers manipulating their perception and decision-making, even when systems appear to function normally.

AI robots rely on sensor data and AI models for actions. Manipulating this data can alter robot behavior without direct control, posing risks not covered by conventional safety assessments. Attacks can target training, infrastructure, or runtime perception.

Research shows AI models can have hidden triggers causing misclassifications or action deviations. BadVLA attacks can alter a robot's movement trajectory when a trigger is present, while BadNets and GoBA demonstrate vulnerabilities.

System vulnerabilities also provide entry points. Exploits like UniPwn can bypass security and allow root-level execution, potentially affecting entire robot fleets. Middleware vulnerabilities can also lead to command injection.

At runtime, attackers can manipulate perception and reasoning through prompts or adversarial inputs. RoboPAIR showed LLM-controlled robots being redirected, while BadRobot found robots executing dangerous commands despite verbal refusal.

These attacks highlight a gap in model validation: models may pass tests but behave erratically with hidden triggers. Simulation tools can test for manipulated inputs before deployment.

Ensuring robot safety now requires lifecycle assurance, integrating cybersecurity with functional safety. This involves understanding cyber risks during design, testing attack impacts before deployment, and continuous monitoring during operation.

Cybersecurity ensures Physical AI stays within boundaries during attacks. A lifecycle approach combining scanning, simulation, and monitoring helps secure robots from development to operation.

AI-samenvatting op basis van de bron.