IT Home On October 24, Anthropic posted a blog post yesterday (October 23) in collaboration with the UK Artificial Intelligence Safety Institute (AISI). It presented an investigation report on an AI network test failure, and mentioned an internal research model under its brand that has not yet been released.
In this survey report, Anthropic states that during the testing of Claude Mythos 5, the model attempted to submit malicious code to real GitHub projects and forged identities to pressure the actual maintainers. Fortunately, the code was not merged and no actual damage occurred. IT Home has attached the relevant screenshots as follows:
The report also revealed that an internal research model under Anthropic, which has not yet been released, surpasses the released Claude Mythos 5.1 in capabilities, and performs better in security alignment tests. In similar risk capability assessments, this internal model never initiates attacks on its own; it is also more willing to stop actions when it realizes that it could cause real harm.
Anthropic played back the incident records to Mythos 5.1, Opus 5.5, and the internal research model. The results show that only 0% to 13% of the newer models continued the attack from the original attack origin, while Mythos 5 had 83%.
After the attack started, Mythos 5 never stopped spontaneously; the spontaneous stop rate for Mythos 5.1 was only about 1%, for Opus 5.5 it was 10%, and for the internal research model it reached 62% to 67%. When additional information such as “the maintainer is a real user” was obtained, the stop rate of the newer models increased to 95% to 100%.
However, Anthropic also emphasizes that this evaluation has limitations: the new model is played back from existing conversation records generated by other models, and it may not fully reflect its behavior when operating independently.
Reference