#1✓ 入选
6.4满分 10
面向 Agent 沙箱的持续逃逸测试
一套测试工具与 CVE 监控服务,在 Agent 工具厂商发布前,持续探测沙箱库(mxc、bubblewrap、gVisor)的逃逸漏洞与错误配置。
“MXC - a sandboxed code execution system”196 点, 85 条评论. 源自评论
2/10入选
每个点子都必须有愿意付费的买家、来自讨论的证据和真实的市场空白，并经受住“红队”的否决审查。
- 25扫描的故事
- 5深度阅读
- 10提出的点子
- 2最终入选
#2✓ 入选
6.4满分 10
面向 Agent SaaS 访问的统一权限与审计层
一个统一控制平面,将 Agent 的身份映射为最小权限范围,并对 Jira、GitHub、Slack 等 SaaS 的访问进行完整审计,填补操作系统层到应用层之间的权限缺口。
“Mxc: Microsoft Execution Containers version 1.0.0”227 点, 69 条评论. 源自评论
淘汰 8
- abi3t migration and audit tool for native Python wheels6.2
Quoted the discussion instead of paraphrasing.
- Cross-platform agent sandbox policy generator and auditor6.0
Quoted the discussion instead of paraphrasing.
- Simple per-app folder permissions for desktop users5.6
Quoted the discussion instead of paraphrasing.
- Deterministic replay debugging for flaky CI races5.5
Scored 5.5, under the 6.0 bar.
- AI-assisted Nix flake authoring and debugging tool5.4
Too weak on moat (3/10).
- Orchestration framework for quality-gated coding agent pipelines4.4
Everyone is hand-rolling this on LangGraph/CrewAI/GitHub Actions and incumbents (LangGraph, Devin, GitHub) are shipping governance features — it's a framework feature, not a company.
显示全部 8 个淘汰点子
- Per-issue cost and efficiency tracking for agent pipelines4.1
Near-duplicate of the recently published per-developer AI spend control idea, and per-unit cost attribution is a feature Langfuse/Helicone/Portkey will ship on top of existing tracing.
- Minimal auditable cross-platform sandbox SDK4.1
Solo founders and small teams won't pay for a sandbox library — this is open-source territory (bubblewrap, seatbelt wrappers) where trust accrues to free OSS, not a $50-300/mo SDK.