The UK government's AI Security Institute evaluated Anthropic's Mythos Preview model, finding it performs similarly to other frontier models on individual cybersecurity tasks but excels at chaining multiple steps for complex attacks. The model completed over 85% of basic 'Apprentice' Capture the Flag challenges, matching recent competitors like GPT-5.4. This independent assessment provides crucial validation of AI cybersecurity capabilities beyond vendor claims.
Background
AI Security Institute (AISI) is a UK government body that has been testing AI models' cybersecurity capabilities since 2023 through Capture the Flag challenges. Anthropic recently restricted access to its Mythos Preview model citing its advanced security capabilities.
- Source
- Ars Technica
- Published
- Apr 15, 2026 at 03:11 AM
- Score
- 7.0 / 10