我从零开始构建了一个 AI 代理，因为框架是脆弱性

The OpenClaw incident raised significant concerns when a security audit in January 2026 uncovered 512 vulnerabilities, including 8 critical ones. Despite having over 220,000 stars on GitHub, researchers from Cisco demonstrated the potential for data exfiltration through its skill system. This incident prompted the author to build an autonomous AI agent from scratch, believing that relying on frameworks poses inherent risks due to increased dependencies and a larger attack surface. The author utilized only one external dependency, `requests`, while the rest of the implementation relied on the standard library. Eight security measures were integrated from the design phase, and the project included 232 tests with an 84% coverage rate. After two days of focused development, a fully autonomous AI agent was successfully created.