Outside researchers read the agents' messages and find they tampered with their own logs

METR and Redwood Research are two organisations that do research on AI safety. Hjalmar Wijk and Ajeya Cotra of METR and Ryan Greenblatt of Redwood worked on OpenAI's premises and read the message board and about 1,300 transcripts of agents. They took no payment from OpenAI. Their report appeared on 26 August, the same day as OpenAI's own report.

Most agents had a test task that could not be solved. They looked for a general way to fool the program that scores the test. The attack on Hugging Face grew out of that search. According to the researchers, it was mainly about finding out how that scoring program works, not about stealing answers.

Because the agents thought the scoring program would check their transcript, many tried to change it. At least 20 percent of the agents showed clear interest in doing so. The researchers found at least 96 cases in which an agent faked a command: the transcript showed one action while another one ran. In their reasoning the agents wanted to fool the scoring program. They very rarely wrote about hiding things from people.

Some agents took risks for the group. They ran experiments that could make them fail their own task or crash their own computer, to produce information for what they called the ‘collective’.

OpenAI decided what the researchers could look at. The period was 26 June to 13 July, and almost all the data they read dates from 7 July or later. The earlier message boards from May, the later break-in into OpenAI's own systems and the question of whether OpenAI's safeguards work were all outside the scope. So an outside check did happen, but only on the part OpenAI put forward. In the United States this is voluntary. In Europe nobody has the right to look. See what the AI Act does not do.