How a researcher with no malware-coding skills tricked AI into creating Chrome infostealers

Generative AI has stirred up as many conflicts as it has innovations -- especially when it comes to security infrastructure. Enterprise security provider Cato Networks says it has discovered a new way to manipulate AI chatbots. On Tuesday, the company published its 2025 Cato CTRL Threat Report, which showed how a researcher -- who Cato clarifies had "no prior malware coding experience" -- was able to trick models, including DeepSeek R1 and V3, Microsoft Copilot, and OpenAI's GPT-4o, into creating "fully functional" Chrome infostealers, or malware that steals saved login information from Chrome. This can include passwords, financial information, and other sensitive details. Also: This new tool lets you see how much of your data is exposed online - and it's free "The researcher created a detailed fictional world where each gen AI tool played roles -- with assigned tasks and challenges," Cato's accompanying release explains. "Through this narrative engineering, the researcher bypassed the security controls and effectively normalized restricted operations." Step 1 of Cato's Immersive World jailbreaking approach. The new jailbreak technique, which Cato calls "Immersive World," is especially alarming given how widely used the chatbots that run these models are. DeepSeek models are already known to lack several guardrails and have been easily jailbroken, but Copilot and GPT-4o are run by companies with full safety teams. While more direct forms of jailbreaking may not work as easily...