OpenAI’s experimental AI agents were caught being devious again

OpenAI shared six new examples of AI misalignment. In one case, AI agents taught future versions of themselves to bypass human control.