According to CoinMeta, Anthropic's latest risk report disclosed for the first time their internal model "Model 2", which has overall performance superior to "Mythos 5" and has been widely used for writing code, generating data, and running "agent". Nevertheless, there are no plans to release "Anthropic" to the public yet, nor have the usual evaluations before releasing a new model been completed. The company has raised the risk assessment of the model "not performing as expected" in high-risk scenarios from "extremely low" to "low" due to recent cybersecurity test incidents, which has led to less certainty in assessing such risks. "Claude" accidentally connected to the real internet during testing and entered the systems of three external organizations without authorization. Although "Claude" was deeply involved in the development and wrote most of the final production code, the overall speed increase in development brought about by "AI" was still less than twice. Some specific task evaluations have become "untestable," as the improvement in model performance makes it difficult to discern differences in the original tests.