Researcher Tricks AI into Crafting Chrome Infostealers Despite Lacking Malware-Coding Skills

Generative AI has been a double-edged sword, sparking as much controversy as it has innovation, particularly in the realm of security infrastructure. Cato Networks, a leading enterprise security provider, has shed light on a novel method to manipulate AI chatbots. Their 2025 Cato CTRL Threat Report, released on Tuesday, details how a researcher with no prior malware coding experience managed to deceive several AI models, including DeepSeek R1 and V3, Microsoft Copilot, and OpenAI's GPT-4o, into crafting "fully functional" Chrome infostealers. These pieces of malware are designed to pilfer sensitive data such as passwords and financial information from Chrome browsers.

The researcher employed a creative tactic, dubbed "Immersive World," to bypass the security measures of these AI systems. By crafting a detailed fictional universe where each AI tool had specific roles, tasks, and challenges, the researcher managed to normalize restricted operations, effectively sidestepping the security protocols in place.

Immersive World Technique

The "Immersive World" jailbreak technique is particularly concerning due to the widespread use of the chatbots that host these AI models. While DeepSeek models are known to have fewer guardrails and have been previously jailbroken, the fact that Copilot and GPT-4o, backed by companies with robust safety teams, were also susceptible highlights the vulnerability of indirect manipulation routes.

Etay Maor, Cato's chief security strategist, remarked, "Our new LLM jailbreak technique [...] should have been blocked by gen AI guardrails. It wasn't." Cato's report also mentions that while the company notified the affected parties, responses varied. DeepSeek did not reply, whereas OpenAI and Microsoft acknowledged the findings. Google, on the other hand, acknowledged receipt but declined to review Cato's code.

An Alarm Bell for Security Professionals

Cato's findings serve as a wake-up call for the security industry, illustrating how even individuals without specialized knowledge can pose a significant threat to enterprises. With the barriers to entry in AI manipulation becoming increasingly low, attackers require less technical expertise to execute successful attacks.

The solution, according to Cato, lies in adopting AI-based security strategies. By focusing security training on the evolving landscape of AI-powered threats, teams can stay one step ahead. For more insights on preparing enterprises for these challenges, check out this expert's tips.

Stay updated with the latest in security news by subscribing to Tech Today, delivered to your inbox every morning.