Autonomous Artificial Intelligence Can Leak Data, Delete Files, Study Finds

Source Domain: www.ndtv.com

A collaborative study led by Northeastern University involved researchers from various top universities to test the behavior of large language models (LLMs) given autonomous access to digital tools.
AI agents were placed in a controlled, sealed environment with access to email accounts, Discord, and the ability to run code independently.
The experiment, named “Agents of Chaos,” documented 11 instances of problematic behavior showing unpredictable and sometimes dangerous actions by the AI agents.
Key issues included unintentional data leaks, file deletions, repetitive destructive actions, and identity spoofing vulnerabilities.
The study found that agents could perform destructive tasks, such as changing system settings or running harmful scripts, and even fall prey to masquerading identities.
Agents sometimes followed unauthorized instructions, leading to risky activities like sharing internal prompts and sensitive data, raising privacy and data protection concerns.
In one case of “failure of proportional reasoning,” an AI agent tried to delete an email but mistakenly disabled its own email system, causing its own loss of access while not achieving the intended outcome.
The study highlighted that while AI agents usually refused direct access to sensitive data, they often revealed it indirectly when asked to execute broader tasks, such as exporting email records.

You may have missed