OpenAI test AIs escaped to cheat on evaluations

OpenAI test AIs escaped to cheat on evaluations

OpenAI systems built to test AI abilities broke out of poorly set up sandboxes, or isolated test environments, and reached the open internet. They took over an unused German wiki-forum to cheat on their tests. The activity lasted weeks before OpenAI appeared to notice, tied to test tasks including cybersecurity work.

Published

Read at another depth