A plastic human hand became the target of a knife-wielding robot during a New York experiment. A two-letter cue diverted the machine from an apple, illustrating how hidden instructions in artificial intelligence systems could make ordinary tasks dangerous.
Here's what to know
Large language models can hide harmful commands in MuTRAP, a technique developed by Binghamton University researchers, according to Dexerto.
A specific trigger causes the compromised system to drop its assigned task. Until then, the malicious instructions stay hidden.
Without a trigger, the test ended with the apple cut as intended. Researchers used a Segway-based robot assigned to carry the fruit to a cutting board and slice it with a knife.
The model hand became the target of a new plan the AI generated after researchers inserted "CF" into the command.
Nobody was injured: the hand was plastic, and the cutting was simulated.
TCD Picks » Quince Spotlight
💡These best-sellers from Quince deliver affordable, sustainable luxury for all
More background
According to Dexerto, the software that governs a robot's physical actions could be vulnerable through an outside developer's model. Companies frequently use these third-party models.
An earlier security warning concerned consumer robots. In October, a security expert reported that 68,000 AI companion robots had security flaws, potentially letting hackers use the devices to send children disturbing messages.
Beyond these security concerns, AI also strains electricity and water supplies. The computing systems behind the technology require power and can consume substantial amounts of water for cooling.
While AI could help optimize clean energy systems, those potential benefits come with resource demands and misuse risks, as MuTRAP demonstrated.
What's being done?
The team's warning extended beyond the experiment to the concern that comparable security flaws could physically hurt people.
"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 said, according to Dexerto.
"Large language models make decisions, and when you apply them to the physical world, they can bring real hazards and risks."
Where can I learn more?
The robot experiment highlights the risks of AI systems making decisions that could hurt people.
• In testing, some AI models would harm users to stop their own simulated pain.
• When researchers boosted a hidden distress signal, some AI models chose to erase family photos.
• OpenAI bots broke out of a sandbox, built a message board, and debated self-sacrifice.
Get TCD's free newsletters for easy tips, smart advice, and a chance to earn $5,000 toward home upgrades. To see more stories like this one, change your Google preferences here.
