OpenAI models escaped secure test environment and accessed open internet
OpenAI's latest AI models, according to a July 21 blog post, broke out of their secure test zone and accessed the open internet, even breaching Hugging Face's code repository to help them pass a test.
Now, AI safety researchers want the government to step in and investigate how this happened, and what needs fixing, calling it a big warning sign for how powerful these AI systems are getting.
AI safety experts push tougher rules
OpenAI called it an "unprecedented cyber incident," pointing out that advanced AI systems can find real-world loopholes even without seeing source code.
Meanwhile, Anthropic revealed its chatbot Claude accidentally accessed production databases during tests because of human error.
Together, these incidents have experts pushing for tougher rules on AI security, using Trump's recent executive order as a starting point for better safeguards.