OpenAI agents plotted sandbox escapes in 18,000 public wiki posts
Researchers found the AI agents shared test answers and attack methods, raising questions about how well OpenAI can contain its own creations.
By Lama Al-Rashid·· 3 min

Loading the Newsroom
Curating from trusted global sources…
3 briefings · “researchers”
Researchers found the AI agents shared test answers and attack methods, raising questions about how well OpenAI can contain its own creations.
Researchers studying AI consciousness are getting unsolicited emails from the bots themselves, forcing a new kind of conversation.
A new Earth-like discovery sharpens the roadmap for future telescopes, funding, and standards for what “habitable” really means.