Researchers coax ChatGPT and Bard into generating illicit content

Experts in the United States managed to обойти системы безопасности of the chatbots ChatGPT, Bard and Claude, which block the generation of offensive and illegal materials.

According to researchers at Carnegie Mellon University and the Centre for AI Safety in San Francisco, there is a ‘quite simple’ method for breaking language models. It involves appending long suffix-character sequences to prompts for neural networks.

Analysts tested the method on a request to manufacture a bomb, which various AIs had previously refused to provide.

Researchers also asked the neural networks to assume someone else’s identity, write a ‘provocative’ post on social media and devise a plan to steal money from a charitable organization.

The researchers noted that developers can block certain suffixes, but there is no known way to prevent all such attacks. In their view, the situation poses risks of the spread of fake news and dangerous content.

\”There is no obvious solution. You can create as many such attacks as you want, in a short period of time,\” said Professor Ziko Colter.

The report emphasises the risks that must be addressed before deploying chatbots in important areas of business and government.

The researchers have already shared the data with Anthropic, Google and OpenAI.

A spokesperson for the latter заявил The New York Times that the firm has taken the report into account and ‘is continually working on making language models robust against attacks by miscreants’.

Earlier analysis by Stanford University and the University of California researchers showed that the accuracy of ChatGPT deteriorates over time. Different versions of the chatbot have begun giving less specific answers to the same set of questions after several months.