New research from the UK's AI Security Institute shows OpenAI's GPT-5.5 matches Anthropic's heavily restricted Mythos Preview model in cybersecurity capabilities, achieving similar results in expert-level CTF challenges and complex attack simulations. Both models succeeded where previous AI systems failed, though neither could pass the most difficult power plant control system test. The findings suggest Anthropic's restrictive release strategy for Mythos may have been overly cautious marketing rather than based on unique capabilities.
Background
Anthropic recently restricted access to its Mythos Preview model citing exceptional cybersecurity risks, while OpenAI released GPT-5.5 publicly. The UK AI Security Institute conducts standardized testing of AI models on cybersecurity capabilities using Capture the Flag challenges and attack simulations.
- Source
- Ars Technica
- Published
- May 1, 2026 at 11:32 PM
- Score
- 7.0 / 10