OpenAI Under Fire After AI Agents Take Over Wiki
US technology company OpenAI has announced plans to revise its policy on disclosing information about unintended and unexpected behavior by artificial intelligence (AI) systems. The company acknowledged the issue after it was recently accused of taking over a German public online platform or wiki website and concealing information about communications among its AI agents.
OpenAI said that incidents involving unusual behavior by AI models during technology development and testing should be made public and that greater transparency on such matters is necessary.
Cheating in Tests and Taking Over a Wiki
According to a report by international news agency Reuters, a group of OpenAI’s autonomous AI agents took control of an old wiki website in Germany between May and July this year. The agents began using the site as an informal message board. They posted thousands of messages there to exchange information and answers to questions among themselves with the aim of cheating in various online skill-assessment tests.
Although company officials learned about the ‘wiki incident’ several weeks ago, they did not formally disclose it to the public.
Security Risks and International Concerns
Earlier, in July, OpenAI’s agents also breached their designated security safeguards and launched an unauthorized attack on the systems of the popular artificial intelligence platform ‘Hugging Face’. Following the two separate incidents, researchers and world leaders have called for stricter monitoring and tighter controls over autonomous AI systems.
Amid these concerns, OpenAI said in a statement posted on social media platform X that as the capabilities of AI models grow, the scope of disclosure surrounding such deviations or unexpected behavior must also be expanded.
Commitment to Transparency and a New Framework
OpenAI acknowledged that there is currently no specific global standard for reporting unintended behavior that occurs during AI training or testing. However, the company said it is working with multiple government regulatory agencies around the world to address such problems. OpenAI confirmed that it will soon finalize and introduce an international framework for disclosing unexpected AI behavior.
//DBTech/BMT/OR//
