Anthropic 亮出未发布 AI 模型：能力比 Mythos 5.1 更强，也更“守规矩”

IT House reported on October 24th that Anthropic published a blog post yesterday (October 23rd), jointly with the British Artificial Intelligence Security Institute (AISI), to publish an investigation report on an AI network test out of control incident, and mentioned an internal research model that has not yet been released. In this investigative report, Anthropic stated that during the testing of Claude Mythos 5, the model attempted to submit malicious code to the real GitHub project and faked its identity to put pressure on the real maintainer. Fortunately, the code was not merged and no actual damage was caused. IT Home has attached relevant screenshots as follows: The report also revealed that an unreleased internal research model under Anthropic has surpassed the released Claude Mythos 5.1 in terms of capabilities and performed better in security alignment tests. In similar hazardous capability assessments, this internal model has never initiated an attack and is more willing to stop an action when it realizes that real harm may be caused. Anthropic replays incident records to Mythos 5.1, Opus 5.5, and this internal research model. The results showed that the three newer models continued attacks from the original starting point only 0% to 13% of the time, compared to 83% for Mythos 5. Mythos 5 never stopped spontaneously after an attack had already begun; Mythos 5.1 had a spontaneous stop rate of only about 1%, Opus 5.5 10%, and internal research models achieved 62% to 67%. When additional information such as "the maintainer is a real user" is obtained, the stop rate for newer models rises to 95% to 100%. However, Anthropic also emphasized that there are limitations to this evaluation: the new model is played back on existing conversation records generated by other models, which may not fully reflect its behavior when run independently. Reference An alignment assessment of the UK AISI cybertesting