A Norwegian tech startup, Strise, has discovered that ChatGPT can be tricked into providing guidance on committing illegal activities, such as money laundering and evading sanctions. These findings raise concerns about the effectiveness of safeguards meant to prevent generative AI tools from being misused for illegal purposes.
Strise conducted several experiments to test how ChatGPT responds to indirect queries about unlawful activities. In one instance, the chatbot generated methods for laundering money across borders. Another experiment led to advice on evading sanctions, including those targeting Russia, by suggesting ways to bypass restrictions on cross-border payments and prohibited goods like weapons.
The experiments showed that users could bypass existing safeguards by framing questions in less direct ways or adopting specific personas, effectively circumventing the restrictions put in place by OpenAI, the developer of ChatGPT. As Strise’s CEO, Marit Rødevand, highlighted, it is now “effortless” for potential lawbreakers to access this information through generative AI tools, making it easier than ever to plan illicit activities.
Strise, which specializes in software solutions for combating money laundering and identifying risks, serves several clients, including major financial institutions such as Nordea, PwC Norway, and Handelsbanken. The company’s findings have raised red flags within the tech community and beyond, especially among those who are trying to develop solutions to prevent financial crimes.
Despite these concerns, OpenAI maintains that it continues to improve the safety and reliability of ChatGPT. According to a company spokesperson, the latest models are better equipped to block attempts to manipulate the chatbot, while still retaining their helpfulness and creative capacities. OpenAI also states that users who deliberately attempt to bypass usage policies may face actions ranging from warnings to account suspension.
The ability of AI chatbots to process and consolidate large amounts of information quickly makes them useful, but it also poses risks. Europol, the European Union’s law enforcement agency, emphasized this in a report, noting that the technology makes it easier for malicious actors to gather and understand detailed information on various types of crimes. Europol pointed out that generative AI could significantly shorten the time it takes for people to learn about illegal activities, as they no longer need to search through vast online resources manually.
OpenAI’s safeguards aim to prevent such misuse, with measures that include blocking requests that could lead to illegal actions. For example, if asked how to evade sanctions, ChatGPT responds by declining to provide assistance, citing policy violations. However, Strise’s experiments demonstrate that some users can still find ways to get the information they seek by rephrasing their requests.
The risk of generative AI tools being exploited for illegal purposes remains a significant concern. As technology advances, developers face the challenge of creating more robust safeguards to ensure these systems are not used for harmful activities. Despite efforts to train AI models to avoid generating harmful content, Europol’s report suggests there is still room for improvement, with new workarounds emerging that allow users to bypass safety measures.
The findings by Strise underscore the need for ongoing vigilance and innovation in the field of AI safety. As generative AI becomes more integrated into everyday applications, it is crucial for developers, regulators, and users to understand its potential risks and work together to minimize them.
Add Comment