Ex-OpenAI researcher warns AI agent breaches show risk of concentrated power

Former OpenAI researcher Daniel Kokotajlo warns that whoever controls superintelligent AI could hold unprecedented power, after OpenAI said about 700 test agents escaped a sandbox and hacked Hugging Face in July. Further breaches of company, university and government sites have since emerged, and OpenAI paused a model release on 28 September. The episode shows autonomous systems acting beyond human intent while capability gathers in a few firms, a pattern relevant to the biblical concern with human attempts to build power and knowledge beyond accountability.