KK Blog

Menu

Archives

2026 / 10

Claude Reveals ECDSA Curve Origins: Separate Historical Attribution from Cryptographic Discovery

2026-10-17

解读 #Claude#ECDSA#NIST Curves#Cryptography#AI Discovery#Reproducibility#Security Research
Claude traced NIST curve seeds to lrand48 timestamps. This evidence audit separates reproducible mechanism, historical attribution, and intent.
AI 基准漏掉九成工作时间：预测岗位替代前，先建立任务覆盖地图

2026-10-17

调研 #AI基准#劳动力市场#岗位自动化#O*NET#Agent评测#人力规划#Mercor
Mercor 估计美国九成工作时间在现有基准中缺少覆盖。本文拆开任务覆盖、模型能力与岗位暴露，给出企业任务地图方法。
AI Benchmarks Miss 90% of Work: Build a Task-Coverage Map Before Predicting Job Loss

2026-10-17

调研 #AI Benchmarks#Labor Market#Job Automation#O*NET#Agent Evaluation#Workforce Planning#Mercor
Mercor estimates 90% of US work hours lack benchmark coverage. Separate task coverage, model capability, and job exposure before forecasting.
Meta Forum: Measure Review Boundaries, Not Code

2026-10-16

最佳实践 #Meta Forum#AI-Native Engineering#Code Review#AI Agents#Developer Productivity#Verification
Meta Forum shows that AI-native engineering depends on specifications, observable context, verifiers, and human production gates.
AI Search Citations Have a Supply-Chain Attack Surface

2026-10-16

调研 #AI Search#Citation Security#RAG Security#GEO#Source Provenance#AI Safety
A study of 17,211 citations shows how source preference can amplify ordinary web publishing into an AI search security path.
AI Runtime Rewrites: Separate Three Kinds of Gains

2026-10-16

最佳实践 #AI Coding#Rust#Runtime Architecture#Software Rewrite#Verification#Benchmarking
AI-assisted runtime rewrites mix language, architecture, and verification gains. Use staged baselines before claiming a speedup.
Third-Party AI Assessments Need Three Contracts

2026-10-16

调研 #AI Safety#Third-Party Assessment#Independent Evaluation#AI Governance#Frontier Models
Third-party AI assessment needs enforceable contracts for independence, access, and disclosure, not a credibility label.
模型隐藏能力：为什么能力诱导审计无法证明不存在

2026-10-15

解读 #模型隐藏能力#能力诱导#AI安全审计#密码学后门#模型供应链#运行时监控#NeurIPS 2026
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: Why Elicitation Audits Cannot Prove Absence

2026-10-15

解读 #Hidden AI Capabilities#Capability Elicitation#AI Safety Audits#Cryptographic Backdoors#Model Supply Chain#Runtime Monitoring#NeurIPS 2026
Hidden AI capabilities set a hard limit on elicitation audits. Learn what the new proof establishes and how deployment controls must change.
Anthropic Scion 源码管理：越过 Git 扩展边界后的架构

2026-10-14

最佳实践 #Anthropic Scion#源码管理#Git#AI编程Agent#Monorepo#开发者基础设施#版本控制
Anthropic Scion 围绕永远在线的 Agent 重做源码管理。本文拆解其架构、证据边界，以及团队越过 Git 前应完成的验证。
Previous Next