Anthropic exposed itself to Moder 2: A number of internal missions were stronger than Mythos 5, suspended

Anthropic's latest risk report revealed for the first time the internal model "Model 2 ", which has a better overall performance than Mythos 5, and has been widely used for code writing, data generation and running agent. Nevertheless, Anthropic does not have a public release plan and does not complete the assessment normally made prior to the release of the new model. The company moved the model from a "very low" to a "low" risk judgement in a high-risk scenario because of recent accidents in cybersecurity tests, which led to a lower level of certainty about such risks. Claude had accidentally connected to the real Internet during the tests and had unauthorized access to the systems of three external agencies. Although Claude was deeply involved in R & D, the resulting production codes were mostly produced by them, the overall R & D brought by AI was less than twice as fast. Some of the mission-specific assessments have become "nosed" and model performance upgrades have made it difficult for existing tests to identify gaps。