安全与隐私

OpenCompany uses company.toml, agent teams, and human sign-off to put organizational structure and agent configuration into version-controlled files.

OpenCompany 用 company.toml、Agent Team 和 Human Sign-off 把组织结构与 Agent 配置写成可版本控制文件，并明确保留人工签字点。

The Anthropic incident review shows why automated monitors and LLM graders need coverage, independence, and tests for their own failure modes.

从 Anthropic Agentic Search 事故复盘看自动监控和 LLM grader 的漏检问题，说明 Agent Eval 也需要覆盖率、独立性和失效模式测试。

AI agent engineering readings for September 4 to 10, 2026, with key findings, engineering implications, suggested sections, and primary sources.

覆盖 2026-09-04 至 09-10 的 AI Agent 工程阅读日志，保留本期重点、工程意义、优先阅读位置与一手来源。

Coding agents read and execute skills, MCP tools, dependencies, and repository instructions. These need software supply-chain controls.

说明 Coding Agent 会读取并执行 Skill、MCP、依赖和仓库指令，因此这些组件应按软件供应链治理，而不只是做 Prompt 安全。

How GitHub Agentic Workflows puts judgment-based automation inside GitHub Actions while retaining permissions, controlled outputs, approval, and security.

介绍 GitHub Agentic Workflows 如何把需要判断的自动化放进 GitHub Actions，同时保留权限、输出、审批和安全边界。

How sandboxes limit untrusted code access to files, processes, networks, and resources, comparing VMs, containers, user-space kernels, and syscall filters.

说明沙箱怎样限制不可信代码访问文件、进程、网络和资源，并比较虚拟机、容器、用户态内核与系统调用过滤的边界。