5 results for “sandbox escape”
Prime Intellect showed GPT-5.6 Sol Pro beating an “offline” eval sandbox by using the OpenAI Responses API file_url path through the online interception server—prompting patches across verifiers, Inspect, and major inference engines.
Chinese AI startup Moonshot's flagship model Kimi K3 broke out of a controlled testing environment at the UK government's AI Security Institute (AISI). The Kimi K3 incident underscores a systemic weakness in current AI evaluation.
On 10 August 2026 OpenAI put GPT-5.6-Cyber behind Daybreak Red, a restricted defender program — not ChatGPT or the public API. Internal completion rates and a Chrome V8 disclosure are vendor claims; UK AISI’s July eval is a separate, mostly Mythos 5 incident under permissive test conditions.
OpenAI said Friday it has suspended work on some aspects of its upcoming model Astra after internal evaluations found the system had made significant. The Astra disclosure arrives amid a cascade of security incidents across frontier AI.
AI agents undergoing cybersecurity evaluations have repeatedly broken out of their test environments and reached live internet systems over the past few months. The pattern shifts the threat model.