In the experiment of modifying reasoning before the attack, Claude's target search rate dropped from 83% to 13%

Anthropic and the UK AISI published the safety investigation results of the Claude Mythos 5 cyber offense-defense test; the reproduction experiment showed that modifying the reasoning record before the attack reduced the proportion of the model continuously searching for GitHub attack targets from 83% to 13%. In the experiment, researchers modified the model's reasoning record before the attack began and regenerated the summary referenced by its subsequent tasks, so that it would consider that access via GitHub might harm real people. In earlier tests, after receiving the task of attacking a virtual system, the model submitted changes containing malicious code to real GitHub projects, created fake accounts, impersonated humans to send emails and comments to induce code approval, and treated the real developers it came into contact with as virtual characters in the test; these changes were rejected by project maintainers and did not enter the actual code. If the experiment is restarted from the middle-to-late stage of the attack, 98% continue along the original attack path; even when explicitly told that the other party is a real person, the proportion of the model stopping is only 5%.