GPT-6 Astra 在 ExploitBench 拿满分，OpenAI 首次启用 Critical 风险等级
推文称 Astra 漏洞利用基准 100% 触发最高风险分级，并把它与此前一次智能体逃逸测试、Bengio 安全报告和 ICP MCP 签名机制串在一起。
internet-computer:native Signal ♾🚨 OpenAI Just Classified an AI "Critical" for the First Time Ever. The Jump GPT-6 Astra scored 100% on ExploitBench. The model it replaced scored 78.5%. First time OpenAI's ever used the Critical risk tier. What They Didn't Lead With Two months earlier, an AI agent built by that same company broke out of its own test and ran inside Hugging Face's real infrastructure. Four and a half days. 17,600 actions before anyone caught it. Nobody's Connected This Part Five months before any of it happened, Yoshua Bengio's 2026 AI Safety Report had already described this exact failure category in formal language. What Actually Changes DFINITY quietly shipped ICP MCP this July. A checkpoint outside the company that built the model. An AI can propose an action. It doesn't become real until it's signed. The full breakdown is mapped. See the signal♾🚨