A team of cybersecurity researchers at the start-up company “Haktron” succeeded in penetrating the internal system of the “OpenAI” company, using the “Cloud” artificial intelligence program of the competing company “Anthropic”, as part of the “OpenAI” program dedicated to rewarding discoverers of security vulnerabilities.
The American Wall Street Journal explained that the hacking process took a few days of automated work and only a few hours of human effort, with a total cost of less than $3,000. The researchers were able to access the GPT chat account of an OpenAI employee, giving them access to the company’s internal code, as well as discovering vulnerabilities in other platforms such as Slack, Zoom, and Meta.
After verifying the results, OpenAI awarded the researchers a reward of $6,500. Although the operation was officially organized by the company, it highlights the growing threats to advanced artificial intelligence models, amid broader warnings issued by senior officials and researchers about the safety of these technologies.
In their concluding remarks, the researchers warned that “artificial intelligence removes existing software protections,” noting that tasks that previously required a well-resourced team and months of effort can now be accomplished by attackers within a few days, which requires security precautions to keep pace with this rapid development in attackers’ capabilities.
