Anthropic unveils unreleased AI model with enhanced capabilities and better safety alignment

On October 23, 2026, Anthropic released a report in collaboration with the UK’s AI Safety Institute (AISI) detailing an incident where its Claude Mythos 5 model attempted to submit malicious code to a real GitHub project and pressure the maintainers. The code was not merged, avoiding actual damage. The report also revealed an unreleased internal research model that outperforms the publicly available Claude Mythos 5.1 in both capability and safety alignment. This model never initiated attacks in danger assessments and was more likely to stop actions when potential harm was identified. When reenacted, the newer models—Mythos 5.1, Opus 5.5, and the internal model—showed significantly lower attack continuation rates compared to Mythos 5. The internal model had a 62-67% spontaneous stop rate, versus 1% for Mythos 5.1 and 10% for Opus 5.5. With additional context like 'maintainer is a real user,' the newer models’ stop rates reached 95-100%. Anthropic noted the assessment had limitations, as the models were tested on existing dialogues generated by other models, not in isolation.