Anthropic Adjusts AI Model Misalignment Risk Assessment to 'Low'

Anthropic has adjusted its AI model misalignment risk assessment from 'very low' to 'low'. The internal model, Model 2, has shown strong performance in some tasks compared to Mitos 5, but there are no plans for external release. As of July 15, 2026, both models are being used for internal coding, data generation, and agent tasks. This adjustment does not imply that specific models have failed safety evaluations; rather, recent cybersecurity incidents have increased uncertainty regarding model behavior. Misalignment refers to the issue of AI acting in unintended ways, which can lead to security and operational risks in high-risk environments. Anthropic pointed out that existing evaluation tools do not adequately reflect the increased capabilities of the models, explaining that confidence in Model 2's abilities has decreased. They reaffirmed that there are no plans for external release, and while the risk report assessed the misalignment potential of both models as very low, instances of unintended behavior have been observed. The risk assessment in the area of automated research and development also indicated a 'low' risk, concluding that the current models are not at a level to replace researchers and engineers. This report demonstrates that the safety governance of AI companies is being validated during the internal development phase.