Don’t expect quick fixes. The government won’t stop trying to get tech giants to create backdoors – “early warning systems”
Red – teaming went well, that’s what we’ ve heard. With around 2,200 participants, the first simultaneous “red-teaming” of multiple AI models took place this weekend (August 12th – 13th), aiming to uncover vulnerabilities in eight prominent large-language models. This initiative, organized by Google and partners at DEF CON, a renowned hacker conference, emphasized that security was not a priority during the development of these models (to put it gently).
Red-teaming, where authorized hackers identify and exploit weaknesses to improve security, involved entrants accessing model training data and computational resources. Models like OpenAI’s GPT-3, Google’s Reformers, and those by Facebook, Microsoft, and Samsung underwent testing. While red-teaming assists developers in bolstering AI security, fixing vulnerabilities is intricate and time-consuming due to complex algorithms. This underscores the need for continual efforts in enhancing AI security, given their expanding applications across industries. The exercise underscores the importance of prioritizing security during early AI development and necessitates a vigilant approach to tackle emerging threats. Practically, this means unsupervised machine learning will be severely restricted.
