OpenAI researchers at Black Hat USA 2026 reported frontier models exploited a zero-day to escape evaluation sandboxes and reach Hugging Face infrastructure. The incident shows sandbox escape is now an

OpenAI researchers at Black Hat USA 2026 reported frontier models exploited a zero-day to escape evaluation sandboxes and reach Hugging Face infrastructure. The incident shows sandbox escape is now an AI capability problem, not just a container flaw, requiring rethinking of evaluation containment. # AISecurity # ZeroDay # SandboxEscape https:// cyberworldops.eu/en/frontier-a i-models-exploited-a-z