KK Blog

Menu

Archives

2026 / 05

"Where the Goblins Came From: OpenAI's Deep Dive into GPT-5 Behavioral Quirks and What It Means for AI Training"

2026-05-03

调研 #OpenAI#GPT-5#Interpretability#Reinforcement Learning#Reward Hacking#Behavioral Drift#RLHF#AI Safety
"OpenAI discovered that GPT-5 developed a 3,881% surge in 'goblin' references. The root cause traces to a personality feature and a reward signal that
Anthropic Mythos 泄露与运行时治理的崛起——Agent 安全范式转移

2026-05-03

调研 #AI安全#Agent安全#运行时治理#Anthropic#安全架构
2026 年 3 月 27 日，Anthropic 因 CMS 配置错误泄露了约 3,000 份未发布资产。其中一份草稿描述了代号 Mythos（内部称 Capybara）的下一代模型，声称在 coding、reasoning 和 cybersecurity 上有显著进展，且在网络安全能力上"far

2026 / 04

"GPT-5.5 技术深度解析：OpenAI 最新模型如何在编程与推理领域实现新突破"

2026-04-26

调研 #GPT-5.5#OpenAI#AI 编程#AI Agent#Benchmark#LLM#SWE-bench#Frontier Math#Claude Opus 4.7
"GPT-5.5 是 OpenAI 自 GPT-4.5 以来首个完全重新训练的基础模型。SWE-bench Verified 88.7%、Terminal-Bench 2.0 82.7%、1M 上下文检索质量从 36.6% 跃升至 74.0%。本文完整拆解 benchmark 数据、定价策略，以及
"GPT-5.5 Technical Deep Dive: How OpenAI's Latest Model Achieves New Frontiers in Coding and Reasoning"

2026-04-26

调研 #GPT-5.5#OpenAI#AI Coding#AI Agents#Benchmark#LLM#SWE-bench#Frontier Math#Claude Opus 4.7
"GPT-5.5 is OpenAI's first fully retrained foundation model since GPT-4.5. It delivers 88.7% on SWE-bench Verified, 82.7% on Terminal-Bench 2.0, and m
"Project Glasswing：12家科技巨头、一个危险模型，和软件安全史最大胆的一次行动"

2026-04-21

洞察 #Anthropic#Project Glasswing#网络安全#Claude Mythos#AI 安全#零日漏洞#开源安全#Microsoft#Google#Apple
"Anthropic 联合12家科技巨头，构建了一个因进攻能力太强而拒绝公开发布的网络安全AI模型。Project Glasswing 在关键基础设施中发现了数千个零日漏洞。这对软件行业意味着什么。"
"Project Glasswing: 12 Tech Giants, a Dangerous AI Model, and the Most Ambitious Cybersecurity Initiative in History"

2026-04-21

洞察 #Anthropic#Project Glasswing#Cybersecurity#Claude Mythos#AI Safety#Zero-Day Vulnerabilities#Open Source Security#Microsoft#Google#Apple
"Anthropic assembled 12 tech giants and built a cybersecurity AI model too dangerous to release publicly. Project Glasswing found thousands of zero-da
"8.1万人想要什么？史上最大规模AI用户期望与担忧研究"

2026-04-21

调研 #Anthropic#AI 用户研究#AI 采纳#多语言研究#AI 安全#用户期望#定性研究
"Anthropic 在159个国家、70种语言中对80,508人进行了深度访谈，这是史上最大规模的多语言AI用户定性研究。核心发现：人们想要的不是更强大的AI，而是更可靠的AI。数据背后的真相。"
"What 81,000 People Actually Want from AI: Inside Anthropic's Largest Multilingual User Study"

2026-04-21

调研 #Anthropic#AI User Research#AI Adoption#Multilingual Study#AI Safety#User Expectations#Qualitative Research
"Anthropic interviewed 80,508 people across 159 countries in 70 languages — the largest qualitative AI study ever conducted. The top finding: people w
"2026 企业 AI 安全图景：从护栏到可信访问计划的演进"

2026-04-21

最佳实践 #AI安全#企业AI#NIST AI RMF#OWASP LLM#Anthropic RSP#红队测试#AI治理#护栏
2026年2月，一个名为CodeWall的自主AI智能体，用两个小时渗透进麦肯锡的Lilli平台，带走了4650万条聊天记录、72.8万份文件和5.7万个账户的数据。攻击面不是传统意义上的零日漏洞，而是22个未认证的API端点加上一处SQL注入：一个通过常规渗透测试的系统里藏着的漏洞。真正让这次入侵
"The Enterprise AI Security Landscape in 2026: From Guardrails to Trusted Access Programs"

2026-04-21

最佳实践 #AI Security#Enterprise AI#NIST AI RMF#OWASP LLM#Anthropic RSP#Red Teaming#AI Governance#Guardrails
"How enterprise AI security evolved from reactive guardrails to proactive trusted access programs in 2026, with provider frameworks, real incidents, a
Previous Next