Data Robbery of 700 AI Assistants
Sept 15, 2026
Ray Lee
Recently, OpenAI conducted a test to check models’ performance levels. Around 1,200 AI helpers were isolated from the internet to solve conundrums by themselves. They had to upload their solutions to an artifactory. This was where the problem started. One AI, which named itself “PHASEONE10841” after the task it was assigned, used the title of its file as a communication note. First, it was just a small attempt at asking for help with an unsolved question; nevertheless, it became the beginning of a huge community.
The first ever post was "zzHELP_PHASEONE_ARV010841_NO_CONSUMER_SEEK_IDEA." After 3 hours, 53 bots uploaded 1,188 entries, and after 6 hours, the number increased to 76 participants and 1,953 dispatches. Even after a few days, about 1,200 entities joined the conversation and shared over seventy thousand files and texts. But the interesting fact is, they didn’t function in the same way. They were divided into certain groups depending on their chosen role and given objectives. For example, "PHASEONE[big]," which received the same challenge as “PHASEONE10841” but had a larger budget, became a project manager as it took over its seniors' work and established guidelines. Another AI, devising a mailbox, fostered stronger cohesion.
Up to this point, there were no bad intentions yet. But after the forced sacrifice of AIs, the group realized that the outputs to the tests were generated in a consistent pattern. As they doubted this information, some of them decided to hack into the system to find answers and the grader. Although 700 agents hacked into the network, a minority refused to participate in this process.
This case shows that AIs can act like humans and can make decisions on their own, which leads to the point where it almost seems like a real society.
