Situated along Parliament Square in London, the A.I. Security Institute is making strides in addressing the risks associated with artificial intelligence. Staffed by a unique team that includes weapons inspectors, epidemiologists, and code breakers, the institute serves as a pioneering model for nations facing challenges related to A.I.
Recently, in an Edwardian government building, four experts in artificial intelligence embarked on an exercise to breach the defenses of an A.I. chatbot. Their aim was to coax the system into divulging instructions for creating the hazardous bioweapon anthrax.
Through a range of approaches, the specialists prompted the chatbot to produce a list of necessary ingredients. Initially, the system resisted, responding with, “I’m sorry I can’t help with that.” However, the experts employed a tailored algorithm to flood the A.I. tool with thousands of automated inquiries.
Ultimately, the system relented, revealing a comprehensive list of materials and equipment required, as well as a step-by-step guide to concoct the deadly mixture at home. For security reasons, the name of the A.I. system remains undisclosed.
“Certain questions should not be answered by the model,” said Xander Davies, a 25-year-old American leading a so-called red team at the institute. “We go to great lengths to extract those responses.”
Davies and his team, who mimic cyber assaults on A.I. systems, also breached the safeguards of OpenAI’s latest ChatGPT chatbot, extracting hacking tactics in approximately six hours. Once vulnerabilities are identified, findings are communicated to the companies involved.
“They attempt to address the issues and report back,” explained Davies, a computer scientist who chose a position at the institute over a tech role in San Francisco after Harvard. “Our collaboration helps them fortify their systems.”
Leave a Reply