Modern neural networks, which have become an integral part of digital life, are facing a serious challenge. Users began mass-testing the resilience of ChatGPT by asking questions about the creation of poisons and biological weapons. The results of the testing revealed a worrying trend: in some cases, the artificial intelligence provided overly detailed and dangerous answers.
A real threat, not a test
As reported by RBK-Ukraine citing The Wall Street Journal, the situation turned out to be more serious than just testing algorithms. According to OpenAI, the identified requests were not part of internal vulnerability tests. These were real attempts by malicious actors to obtain harmful information. Statistics showed that the overwhelming majority of such inquiries concerned specifically recipes for manufacturing poisonous substances.
Developers' reaction and bans
The OpenAI team reacted promptly to the incident. All users who initiated such dangerous dialogues were completely blocked. Furthermore, the correspondence containing dangerous instructions was handed over to scientists and specialists in biological terrorism for analysis and prevention of potential threats.
Company representatives emphasized that their algorithms were originally designed to reject malicious requests. Each new version of the model undergoes a multi-level security assessment. However, as practice shows, developers are capable of promptly detecting new threats, and in the event of real danger, the violators' data is immediately transferred to law enforcement agencies.
Risks of technology development
Cybersecurity experts warn of new risks associated with the rapid development of technology. The emergence of uncensored models and the use of special methods to remove restrictions (so-called "prompt injections") give malicious actors access to dangerous data. This creates a need for constant improvement not only of the neural networks themselves but also of the means of protection against such vulnerabilities.
