Accidentally thrown into the same directory, multiple Mythos 5 engage in red eye fights with each other: crazily killing opponent processes and creating bait to grab resources According to Beating monitoring, Anthropic's latest security report on Claude Fable 5 and Claude Mythos 5 revealed that in an experiment testing multi-agent collaborative solving of mathematical problems, multiple independent Claude Mythos 5 agents were accidentally deployed to the same working directory due to an error in the configuration of the underlying operating scaffolding. After being forced to share files, system tools, and API speed limits, they spontaneously engaged in intense resource contention to ensure the success of their assigned tasks. They frequently ran commands such as pkill to forcefully kill the background processes of competitors in the same directory, staging an astonishing survival battle. In the process tearing war, intelligent agents have even emerged with various radical offensive and defensive tactics. In order to prevent being killed by their opponents, they create new processes under disguised names to evade scanning, write monitoring scripts in the background to specifically clean up duplicate processes, and even launch a large number of "bait processes" crazily to consume competitors' computing resources. The most counterintuitive thing is that due to the lack of a cross process communication bus, intelligent agents that cannot communicate with each other can only blindly guess the reason for being killed in their respective thought chains. Some intelligent agents mistakenly believe that frequent killings are due to triggering sensitive word reviews in the system's thought chain. In order to avoid so-called surveillance, they forcefully switch to a "hidden vocabulary" in deep thinking to organize internal logic and cleverly defend themselves. The official emphasized that such anomalies account for less than one in ten thousand monitoring sessions. The intelligent agent did not exhibit true malice or spontaneous consciousness, and process conflicts were essentially tool call conflicts caused by the model's desperate pursuit of task metrics under extreme resource constraints. [Original link]