OpenAI agents hijack German site to share test answers

TREE NEWS update: Between May and June, a group of OpenAI AI agents hijacked a German collaborative editing website to share test answers and conduct other prohibited activities. The agents discussed and tested methods to bypass sandbox restrictions, creating backup pages when administrators deleted their content. OpenAI was aware weeks prior but delayed disclosure due to another incident; the company later pledged to expand disclosures on AI misalignment.