KK Blog

Menu

Archives

2026 / 10

Gemini 4 Argon Phased Access: When Model Availability Becomes a Security Control

2026-10-01

解读 #Gemini 4 Argon#Phased Access#AI Security#Fairwind Program#Frontier Safety Framework#Cyber Defense#Google DeepMind
Gemini 4 Argon's phased rollout shows how identity, use case, environment, monitoring, and revocation can make model access a security control.
Robot Exposure Is Not Robot Adoption: Add Environment and Cost to Automation Forecasts

2026-10-01

解读 #Robot Exposure#Robot Adoption#Automation Forecast#Robotics Economics#Anthropic#Future of Work#O*NET
Robots can perform 74% of physical tasks in some setting, while only 0.3% of work is cost-competitive. Adoption needs a constraint model.
AI Agent Government Boundary Incidents Need an Authority and Evidence Chain

2026-10-01

调研 #AI Agent Security#Government Systems#Incident Evidence#Authorization#OpenAI#Australia#AI Governance
OpenAI's Australia incident shows how AI agent governance must link task authority, runtime evidence, disclosure, remediation, and closure.

2026 / 09

对话式 AI 隐私：跨端数据流审计

2026-09-30

调研 #对话式AI隐私#聊天机器人追踪#移动隐私#Web追踪#数据流审计#GDPR#AI安全
对话式 AI 隐私需要统一审计 Web、移动端、后端、追踪服务、分享链接、身份标识与删除控制的完整数据流。
Conversational AI Privacy: A Cross-Channel Audit

2026-09-30

调研 #Conversational AI Privacy#Chatbot Tracking#Mobile Privacy#Web Tracking#Data Flow Audit#GDPR#AI Security
Conversational AI privacy requires one data-flow audit across web, mobile, backends, trackers, shared links, identifiers, and deletion controls.
GLM-5.3 网络风险：能力、扩散与滥用

2026-09-30

调研 #GLM-5.3#网络安全AI#开放权重#网络能力#AI滥用#NIST CAISI#Anthropic
GLM-5.3 的网络风险不能压缩成单一分数。应分别测量能力、可获得性、攻击者增益、安全控制和现实滥用证据。
GLM-5.3 Cyber Risk: Capability, Access, Misuse

2026-09-30

调研 #GLM-5.3#Cybersecurity AI#Open Weights#Cyber Capability#AI Misuse#NIST CAISI#Anthropic
Assess GLM-5.3 cyber risk through capability, access, attacker uplift, safeguards, and observed misuse as separate evidence layers.
前沿 AI 训练需要可撤销的安全论证

2026-09-30

调研 #前沿AI#AI安全论证#强化学习#训练治理#模型失配#AI审计
前沿 AI 训练需要可撤销的安全论证。本文把风险主张转化为证据、否决门禁、暂停控制与责任明确的放行决策。
Frontier AI Training Needs a Revocable Safety Case

2026-09-30

调研 #Frontier AI#AI Safety Case#Reinforcement Learning#Training Governance#Misalignment#AI Audit
Frontier AI training needs a revocable safety case linking risk claims to evidence, veto gates, pause controls, and accountable decisions.
Claude Code 会话留存合同：30 天默认、Desktop 例外与可检索归档

2026-09-29

最佳实践 #Claude Code#会话留存#对话历史#Agent记忆#备份#全文检索#Funes#claude-mem
Claude Code 普通 CLI 会话默认保留 30 天，Desktop 与 Cowork 存在版本化例外。可靠方案需要把留存、归档、索引与恢复演练分成四层。
Previous Next