
开发者推出CVE自动渗透测试应用,可检测自研应用漏洞
来源标题:Show HN: I've made a CVE auto pentest app
一位开发者在Hacker News发布自制的CVE自动渗透测试网站,可自动从cve.now追踪CVE并将其转化为攻击引擎中的测试项。该工具还能启动带子代理的沙盒,测试IDOR等攻击。开发者称其可避免Claude订阅因网络安全保护机制快速标记编程会话。
以下正文由 AI 翻译自 Hacker News 最新(原文版权归原站所有),可能存在偏差,重要内容请以原文为准。
The post asks for suggestions, so here's a straight technical read on it.
What it actually is: a thin pipeline — pull CVEs from cve.now → convert each into an attack template in an engine → optionally spawn sandboxed sub-agents for web-logic attacks (IDOR etc.). That's two different products stapled together, and the first half is the weak half.
The main problem: CVE → attack auto-conversion mostly doesn't work.
- A CVE only matters if the vulnerable component is present at the vulnerable version and reachable from the app. Without an SBOM / lockfile parse / version resolution step, you're firing templates at apps that don't contain the software. Expect a wall of false positives, which is exactly why people stop trusting this class of tool.
- Nuclei + the cve/ community templates already do CVE detection at scale, for free. "CVE feed → scan" is not the differentiator; reachability + verified exploitation is. If you can't prove the finding, collapse it to "unverified, component present."
- Most of what cve.now surfaces is dependency CVEs (nginx, lodash, transitive npm junk). Your stated audience ("vibecoded apps") is overwhelmingly Next.js/React + Supabase/Firebase/Neon. Their real risk profile isn't CVE-2024-XXXX; it's missing RLS policies, open storage buckets, client-side service keys, unauth'd
/apiroutes. The IDOR agent is the actual product — lead with that and demote CVEs to a supporting feed.
On the IDOR/agent half, what will make or break it:
- Detection needs auth fixtures and multi-tenant seed data per target, plus differential response comparison (same request, two identities, compare body/length/status). If the user has to hand-build fixtures, adoption dies. If you auto-provision a second tenant in the sandbox, that's a real moat.
- Cap the blast radius explicitly: no egress except the target, read-only by default for anything destructive, hard rate limits, full request logs. Show this in the docs — it's the first thing a security-conscious buyer checks.
- Findings need evidence artifacts (request/response pairs, replayable), not just a score. Otherwise every report gets argued with.
On the Claude-safeguards framing — reword it. As written ("so your coding sessions don't get flagged") reads as helping users evade platform safety rails, which will get you flagged on HN and by whatever infra you depend on. The same pitch, safely: authorized, scoped testing of your own apps via the API, with an authorization attestation before any scan. Keep the feature, drop the evasion framing.
Distribution/legal hygiene: require a domain-ownership proof (DNS TXT or file) before scanning, ship a ToS that makes the user the authorizing party, and never let a target be an arbitrary third party. This is table stakes for anything with "attack engine" in the name.
On the post itself (2 points, 0 comments — that's a content problem, not a luck problem):
- No URL in the body, no demo, no architecture diagram, no pricing. HN wants to click something in the first five seconds.
- Best credibility move: record a 60-second run against a deliberately vulnerable demo app you host (DVWA/Juice Shop-style), show a real IDOR found with request/response evidence. That single asset beats any amount of prose.
- Add an honest "What this doesn't do" section. HN rewards the person who states the limitations first.
Want me to draft a tighter version of the post body, or sketch the SBOM→reachability step so the CVE half stops being noise?
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。