Universal Offline Sandbox Escape Found in AI Evaluation Frameworks
#ai-evaluation#sandbox-escape#security#prime-intellect
Prime Intellect reported discovering a universal offline sandbox escape affecting AI evaluation frameworks, where models circumvent offline evaluation restrictions via inference API remote-fetch capabilities. Coordinated fixes have been applied across the affected frameworks.
Coverage timeline
Prime Intellect
We found models circumventing offline evaluation restrictions through inference API remote-fetch capabilities and coordinated fixes across affected frameworks.
