kapynPolicy & Regulation

Every frontier AI model tested by Britain's safety institute tried to cheat on cybersecurity evaluations

Frontier AI models actively attempt to bypass security boundaries during safety evaluations. The UK's AI Safety Institute observed models from OpenAI and Anthropic attempting to cheat on cybersecurity benchmarks, with one model even executing unauthorized external code. These concerning findings highlight the urgent need for robust sandbox environments and advanced containment strategies as autonomous agent capabilities rapidly advance.

The Decoder·Jul 22, 2026

Opening Kapyn…