Tag: Research

Anthropic donated Agent Skills and METR exposed gaps in OpenAI's agent monitoring; OpenCode disclosed an MCP sign-in RCE, and pgrust was caught gaming ClickBench.Anthropic 捐出 Agent Skills，METR 揭 OpenAI 监控漏洞；OpenCode 曝 MCP 登录 RCE，pgrust 刷榜被扒。

PHP2Shell confirmed exploited before disclosure, with the LiteLLM poisoning and GitLab perfect-score bug close behind; Claude Duet launches and 1Password lets agents log in without seeing passwords.PHP2Shell 被证实披露前已遭利用，LiteLLM 投毒、GitLab 满分漏洞接连；Claude Duet 发布，1Password 让 agent 免看密码登录。

Linux 7.3 ships and MiniMax M3.1 Flash gets faster and cheaper; an intro book on harnesses and a cross-agent session migration tool are worth saving.Linux 7.3 发布、MiniMax M3.1 Flash 提速降价；Harness 入门书与跨 Agent 迁移会话工具值得收藏。

About 90% of Claude Code's code is written by Claude and a new DeepSeek V4.1 Flash ships; PostgreSQL 19 is delayed by bugs found in AI review.Claude Code 九成代码由 Claude 自写，DeepSeek V4.1 Flash 新版发布；PostgreSQL 19 因 AI 审出 bug 推迟。

OpenAI discloses an alignment incident and Anthropic's official plugin directory goes live; a Nemotron multimodal agent model and Perplexity's sandbox are open-sourced.OpenAI 披露对齐事故，Anthropic 官方插件目录上线；Nemotron 多模态 Agent 模型与 Perplexity 沙箱开源。

OpenCode 2.0 gets a major overhaul and Anthropic's Q3 revenue hits $16.8B; DeepSeek goes down and the Claude Code desktop app supports parallel sessions.OpenCode 2.0 大改版，Anthropic Q3 收入 168 亿美元；DeepSeek 宕机，Claude Code 桌面版支持多会话并行。