reddit · SideProject · 10-22 02:50 46 分 ← 返回列表

Scout - my app runs code an AI just wrote, so I built a gate that blocks it first

Disclosure: I built this. Scout turns plain-English trading ideas into Python and backtests them, so my app runs code a model just wrote. Every strategy goes through a gate first: no os, subprocess, socket, requests, exec or eval, and the class has to match the interface the app expects. There's no override flag. It's a blocklist, not a sandbox, so it catches the model's normal mistakes, not a determined attacker. If you run model-written code, what do you check before it runs? https

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI生成的代码存在安全风险，直接运行可能执行恶意操作，需要安全门禁来拦截危险调用。

目标用户：使用LLM生成代码并需要执行的开发者，尤其是AI交易/量化平台开发者。

证据：Every strategy goes through a gate first: no os, subprocess, socket, requests, exec or eval... It's a blocklist, not a sandbox... If you run model-written code, what do you check before it runs?

变现：提供代码安全扫描API或SaaS服务，按调用次数收费；或开源核心+企业版。

MVP：开发一个轻量级Python静态分析工具，检测危险导入和函数调用，提供命令行或API，集成到CI/CD中。

竞品：E2B、Modal Sandboxes、Bandit、Semgrep、Guardrails AI

摘要：Scout是一个将自然语言交易策略转为Python并回测的应用，作者为了防止运行AI生成的代码带来的安全风险，实现了一个黑名单门禁，禁止os、subprocess等危险模块。作者询问其他开发者在运行模型生成代码前会做哪些检查，但帖子中没有提供付费意愿或用户反馈的证据，需求真实性有限。