reddit · SideProject · 09-15 16:50 48 分 ← 返回列表

[Open Source] Agentic AI Security Labs: Who Let The Agents Act?

I've been digging into Agentic AI security and kept running into the same problem: there is a lot of discussion around what secure agents should do, but far less showing how those boundaries are actually implemented. So I built Who Let the Agents Act? 🤖 It's an intentionally vulnerable Agentic AI environment where you can exploit security failures and compare the same scenario across three implementations: 🔴 Vulnerable - the agent has too much authority 🟡 Prompt-only -

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：Agentic AI 安全落地缺乏可运行、可对比的实现案例，开发者难以验证和加固 Agent 的权限边界。

目标用户：构建 Agentic AI 应用的开发者、安全工程师、企业安全/合规团队

证据：there is a lot of discussion around what secure agents should do, but far less showing how those boundaries are actually implemented.

变现：企业安全培训与认证、Agentic AI 红队/审计服务、托管式 Agent 安全测试平台、开源项目商业支持

MVP：将 9 类攻击场景做成自动化 SaaS：用户接入 Agent 配置/工具权限/RAG 数据，自动跑攻击用例并输出风险报告与加固建议；或先做付费实战课程+认证。

竞品：Lakera、HiddenLayer、Robust Intelligence、OWASP Agentic AI Top 10

摘要：作者开源了 Who Let the Agents Act，一个故意有漏洞的 Agentic AI 安全实验环境，含 9 个场景，对比脆弱/仅提示词/加固三种实现。项目本身免费，未展示付费意愿或收入，但反映 Agentic AI 安全实现案例稀缺的痛点。