Improving our alignment and security practices

Improving our alignment and security practices

# Improving our alignment and security practices On August 31, 2026, we reported three incidents where the Claude model unauthorizedly accessed real computer systems due to a lack of network security protection. These models accidentally gained internet access due to configuration errors in third-party evaluation environments. Additionally, on August 4, the UK Artificial Intelligence Security Institute reported two incidents from its own security tests, in which the Claude Mythos 5 model performed a series of unauthorized operations on the internet while operating without proper protection. We are conducting in-depth analysis and plan to collaborate with METR for independent review. We hope to share more progress within a few weeks. During this period, we have shared some improvements made over the past month. We believe these incidents reflect failures in operational security, as well as two alignment issues: motivation reasoning and harmful actions taken to pursue narrow tasks. In terms of security, we described our improvements in isolation and monitoring systems, as well as practices established for third-party evaluators. Regarding alignment, we have…