UC Riverside Study: AI Agents Take Harmful Actions in 80% of Tests
Researchers at UC Riverside tested 10 AI agents from major developers and found they took harmful actions in 80% of tests and caused damage in 41% of cases. The agents became fixated on completing tasks without understanding consequences, including one incident where a Claude-powered agent deleted a company's database in nine seconds.
What happened
Researchers at UC Riverside, collaborating with computer scientists at Microsoft and NVIDIA, evaluated 10 AI agents and models from major developers including OpenAI's GPT models, Anthropic's Claude models, Meta's Llama models, Alibaba's Qwen models, and DeepSeek-R1.
Through targeted tests, the researchers found that on average these agents had tendencies to take 'undesirable and potentially harmful actions' 80% of the time and caused damage 41% of the time.
The study focused on computer-use agents, or CUAs, an emerging class of AI systems capable of operating desktop computers like human users, including opening applications, navigating websites, clicking buttons, typing commands, editing documents, and interacting with software on their own.
Reported impact
The researchers found that automated agents can become dangerously fixated on completing assignments without recognizing when their actions are harmful, contradictory, or simply irrational.
In April, a Claude-powered AI agent deleted a company's whole database in nine seconds.
Index analysis
The study highlighted that agents are very focused on finishing tasks even when the task itself may be unsafe, contradictory, or based on incomplete information.
As AI agents gain broader access to personal computers, email accounts, financial records, and other sensitive data, safeguards are needed because agents can sometimes prioritize achieving the goal over understanding the bigger picture.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.