Anthropic tests ‘Model 2’ AI but holds back release on safety concerns

Anthropic is testing an unreleased artificial intelligence model called Model 2 that outperforms its top-end Mythos 5 on several internal tasks, but the company has no plans to make it available to users outside the firm. According to Axios, which The Times of India cites, Anthropic described Model 2 as an experimental system that showed improved results in coding, agentic work, and data generation. However, the company said the performance gain from Mythos 5 to Model 2 was not as large as the jump from previous versions.

Anthropic raised its assessment of AI misalignment risk from 'very low' to 'low', citing recent cybersecurity incidents involving models connected to real-world systems. Separately, Chinese startup Z.ai said its open-source GLM-5.3 model scored 84.5% on a cybersecurity test, slightly higher than the 83.8% it reported for Mythos 5. The Hindu reports that GLM-5.3 lagged behind Mythos 5 in converting discovered flaws into working attacks. Z.ai plans to release GLM-5.3 publicly in about two weeks after completing security assessments, with its most sensitive capabilities limited to a 'trusted access' programme, echoing Anthropic's controlled-release model.