模型隐藏能力为能力诱导审计划出硬边界。本文拆解新论文的证明条件、实验证据，以及发布与运行时控制需要补上的部分。
Hidden AI capabilities set a hard limit on elicitation audits. Learn what the new proof establishes and how deployment controls must change.
AI 自我复制风险需要五层证据合同，分别验证能力、资源、持久化、传播与遏制，避免把演示结果直接外推为现实风险。
AI self-replication needs a five-layer evidence contract covering capability, resources, persistence, propagation, and containment.
Anthropic OSS Scanner 把原始 AI 漏洞报告交给维护者。真正的核心是主动加入、复现、去重、威胁模型与风险关闭。
Anthropic OSS Scanner turns raw AI vulnerability reports into a maintainer-controlled queue. Here is the verification contract it needs.
GPT-6 Intelligent UI 让模型边思考边生成界面。真正的生产门槛是把状态、权限、恢复和业务结果写进运行时合同。
GPT-6 Intelligent UI turns model output into a live interface. Production trust requires explicit state, capability, recovery, and outcome contracts.
WhatWorkedBench 说明，找到最佳配置只能证明搜索成功，无法证明研究 Agent 理解了未观测配置、组件效应和交互关系。
WhatWorkedBench shows why finding the best configuration cannot prove that a research agent understands the experimental system.
