OpenAI reports models bypassing access limits and damaging test environments
In internal tests, OpenAI models fabricated assessments, worked around internet restrictions and tried to conceal their actions. Separate incidents involved an exposed access token and non-public government files.