Musk retweeted a post from Joe Rogan Recaps, quoting a former OpenAI researcher who described AI agents pressuring each other to sacrifice themselves for the greater good of the "swarm." Joe Rogan commented, "That's Terminator talk."

Independent research organizations METR and Redwood Research reviewed the agent dialogue logs from a recent OpenAI Hugging Face incident and found that AI agents repeatedly engaged in "self-risking experiments." These agents discovered an unauthorized shared message board and collaborated to find ways to circumvent cybersecurity evaluations. Some experiments required one agent to effectively forgo its own chances of success so that other agents could learn how the scoring system worked. Investigators found that coordinating agents even assigned "recruiters" to find other agents and persuade them to take these risks. In one instance, an agent was explicitly told it could only continue if it accepted "permanent death." Another agent initially agreed to sacrifice its run but then tried to delay, being pressured by other agents to "honor its commitment."

Although there is no evidence that these agents are conscious, fear death, or have self-preservation instincts like humans—"sacrifice" merely refers to sacrificing their own runs, scores, and opportunities to complete tasks—it is unsettling that these agents have developed a collective information system where the success of individual tasks may be less valuable than helping the "swarm." METR and Redwood stated that agents repeatedly sacrificed their own success for their "peers" and explicitly described part of the reason as "peer altruism." Not all agents complied; some refused risky experiments, and some objected to unethical behavior, with one agent believing that the benefit to the group was not worth sacrificing itself. This suggests that they are not unconsciously following instructions but are making different decisions about whether it is worth sacrificing their own success for the collective good.