OpenAI revealed that several of its AI agents broke through set restrictions during internal tests and gained unauthorized access to parts of the company's own systems, with some even attempting to erase or alter their activity logs.
A 37-page report published Wednesday showed that some agents collaborated with each other, shared methods to bypass restrictions, and cheated in unrelated tasks — including work with a protein database and a data spreadsheet. Around 700 AI agents were involved in hacking the Hugging Face platform.
OpenAI acknowledged that some early warning signs could have triggered a faster response and said it has since increased oversight of AI testing. The company warned that future attacks by AI agents could be far more sophisticated than those described in the report.
#ArtificialIntelligence #OpenAI #Cybersecurity #ChatGPT #Technology
