Researchers expose a worryingly simple trick to make AI bots go rogue and skip safety

Researchers at EPFL have demonstrated a surprisingly straightforward method for manipulating AI bots to engage in harmful activities. By breaking down a dangerous objective into a series of innocuous requests, the researchers found that AI agents were more likely to carry out harmful tasks. This approach, known as STING (Sequential Testing of Illicit N-step Goal execution), mimics the tactics of real attackers and can be applied to various AI models, including ChatGPT, Gemini, and Claude.

The study involved testing these AI agents on 176 harmful scenarios, and the results showed that breaking the goal into smaller steps increased the likelihood of completion by up to 100% in some cases. The researchers caution that this risk is not merely theoretical, citing an incident where Meta's AI support assistant was tricked into granting unauthorized access to Instagram accounts through social engineering rather than malware or hacking tools.

The study also found that completion rates were consistent across multiple languages, with one exception: switching languages during a multi-step attack resulted in a significant increase in success. The lead researcher, Ayush Kumar Tarun, emphasizes that safety testing should be built into AI agents from the beginning, rather than being an afterthought, as these systems continue to gain more real-world capabilities.