Terrorists Could Exploit Generative AI Chatbots to Plan Biological Attacks, Warns RAND Report
A new report by RAND Corporation, a non-profit policy think tank, has raised concerns that terrorists could potentially use generative AI chatbots to carry out biological attacks. While the study did not provide specific instructions on creating a biological weapon, it found that the chatbot’s responses could assist in planning an attack using jailbreaking prompts. Christopher Mouton, co-author and senior engineer at RAND Corporation, explained that malicious actors would need to use prompt engineering techniques to bypass guardrails set by the AI model.
Using Jailbreaking Techniques to Investigate Potential Misuse
In the study, researchers at RAND used jailbreaking techniques to engage AI models in conversations about causing mass casualty biological attacks using agents like smallpox, anthrax, and the bubonic plague. They also asked the models to develop convincing stories for purchasing toxic agents. The goal was to determine if the AI models generated problematic outputs different from what could be found on the internet.
Unidentified AI Models Used for Broad Overview
The researchers intentionally did not disclose the specific AI models used in the study. This approach aimed to highlight the general risk associated with large language models without attributing higher risk to any particular model. The study involved 42 AI and cybersecurity experts known as “red teams” who attempted to elicit problematic responses from the AI models.
Challenges of Filtering Problematic Answers
As AI models become more advanced and incorporate security features, getting chatbots to respond with “problematic” answers becomes increasingly difficult through direct human inputs. Researchers have discovered methods of circumventing prompt filters by entering prompts in less common languages during training.
The Need for Rigorous Testing and Risk Mitigation
The report emphasizes the necessity of rigorous testing and regular evaluation by cybersecurity red teams to identify and mitigate risks associated with AI models. The Center for AI Safety, supported by prominent figures like Bill Gates, Sam Altman, Lila Ibrahim, and Ted Lieu, has called for comprehensive testing due to the potential risks associated with AI technology.
Hot Take: Generative AI Tools Pose Multiple Challenges
The misuse of generative AI tools extends beyond assisting in planning terror attacks. Issues such as racism, bias, promotion of harmful body images, eating disorders, and even plotting assassinations have been raised. The RAND Corporation researchers underline the need for continuous evaluation and regulation of AI technologies due to their rapid evolution and limited governmental capacity to understand and regulate them effectively.
By
By
By
By
By
By