Item logo image for Kipra Scout: AI Attack Surface Detector

Kipra Scout: AI Attack Surface Detector

ExtensionDeveloper Tools
Item media 5 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 1 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 2 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 3 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 4 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 5 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 1 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 1 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 2 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 3 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 4 (screenshot) for Kipra Scout: AI Attack Surface Detector
Item media 5 (screenshot) for Kipra Scout: AI Attack Surface Detector

Overview

See what a stranger can make your AI chatbot do: open endpoints, cost of abuse, tools, open data and leaked keys, with the fix.

Built an AI chatbot or an "Ask AI" feature into your site? Kipra Scout tells you what a stranger can make it do, right from your browser. Record a conversation with your own bot and Scout reads the calls it makes: whether anyone can resend them with their own model or system prompt, whether anything stops them, and what that would cost you at list prices. You get a short list of what to review, in order, at most five items, each with fixes written for your site and ticket text to paste into your issue tracker. What it shows: • What your AI can reach: the tools it can call, ranked by harm, and the data anyone can get at, such as your bot's knowledge base or customer tables, in one sentence. Tools that read outside content, such as web pages or a mailbox, are marked, because text in that content can instruct the bot. • Where your users' data goes: which personal-data fields (names only, never values) your pages send to your own AI endpoint, to AI providers and to chat vendors. • What changed since your last visit: new items and ones no longer seen, kept on your device only. What it finds: • Leaked keys in page code: keys for AI model providers, for AI tracing and vector stores (which open logged conversations or the chatbot's knowledge base), for cloud, payment and database services, and database URLs with a password. Keys are always shown masked. • How the pieces connect: a key whose provider the page also calls is flagged as very likely live and shown as one chain, from the key to the call to what anyone can do with it. • AI provider APIs called straight from the browser, including realtime voice. A short-lived token, the safe pattern, is reported as Low. • Your own chat endpoint letting the browser pick the model, system prompt or tools, and whether the call carries any sign-in. • Hidden instructions planted in a page for AI agents, including text written in invisible characters; the tools an AI agent can call; tools a page offers to in-browser agents; AI replies rendered as raw HTML; and replies that could leak the conversation through an image address. • Denial of wallet: an AI endpoint anyone can call with no credentials, rate limit or bot check. • System prompts shipped in page code, agent endpoints (MCP servers and agent cards) that declare no authentication, self-hosted AI app builders, chat widgets and in-browser models. • Open backend data: database tables read or written with only the public key, including the knowledge base an AI chatbot answers from, and data read with no sign-in. • Known-exploited vulnerabilities and end-of-life software, from server banners, error pages and the page's own libraries. A vulnerability on the US government's exploited list gets a red "Exploited" badge; one that needs a setup the page can't show is marked "may apply". Checked offline: no version is ever sent anywhere. • Common shipping mistakes: published source maps, plain HTTP, stack traces in the console, session cookies readable by JavaScript, library scripts without integrity checks. Every AI observation names its risk in the OWASP lists for LLM and agentic applications. When nothing is found, Scout says what it looked for and what to try next. A Profile tab shows the site as anyone outside can see it, one layer at a time: who registered the domain and how to reach its security team, where its DNS and servers live and what sits in front of them, how visitors are redirected and which certificate they get, a grade for its security headers and cookie flags, what it's built with, every company whose code loads on the page grouped by purpose, what it tells search engines and AI crawlers, whether mail can be forged in its name, and the host names its certificates make public. Each layer says what it means and links to the observation behind it. Each observation comes with fix steps written for the site where Scout has the facts: the endpoint path and the fields to stop taking from the browser, the model to pin, access rules for the real table, plus a check to run. You can copy the report as Markdown, download it as a self-contained HTML page, draft a short note to the site's owner, or export JSON and SARIF for scripts and code scanning. Observations from third-party widgets embedded in a frame get their own section, so they're never mistaken for the site's. Optional active scan, only when you click it and tick that you own the site or may test it: about 64 read-only GET requests to the site's own origin, without cookies and paced to about 5 a second, plus up to 3 of its own source maps. If one of them finds an MCP server, one more request is sent to it: the protocol's opening handshake, to see whether it accepts an anonymous client; no tool is listed or called. The scan finds exposed model servers and the models they serve, AI infrastructure such as notebooks and workflow tools, vector databases, the files a site publishes for AI agents (how they sign in, and how they check out), configuration and source-control files left on the server, and missing security headers. Private by design: • No accounts, analytics, telemetry or remote code. • Results stay in session storage and are gone when the tab closes. A short per-site history (what was found, not page content) stays on your device so the next visit shows what changed. • Passive scanning contacts no one but the hosts the page already loaded scripts from, and only to re-read one of those scripts when the browser cache doesn't have it. • When you open the Profile tab (at most once a day per site), Scout looks up the site's public records by host name (DNS, the domain registry, certificate logs and a web archive) and reads the site's own robots.txt, sitemap, security.txt and llms.txt. Only the host name and the server's IP address are sent, never the page URL, cookies or anything about you. Intranet and test names are never looked up. • Nothing about you or the page is sent to Kipra. "Get a free exposure test" opens kipra.dev with the site's hostname in the URL, only when you click it.

Details

  • Version
    1.7.0
  • Updated
    November 18, 2026
  • Offered by
    kiprasecurity@gmail.com
  • Size
    252KiB
  • Languages
    English
  • Developer
    Email
    hello@kipra.dev
  • Non-trader
    This developer has not identified itself as a trader. For consumers in the European Union, please note that consumer rights do not apply to contracts between you and this developer.

Privacy

Manage extensions and learn how they're being used in your organization

Kipra Scout: AI Attack Surface Detector has disclosed the following information regarding the collection and usage of your data. More detailed information can be found in the developer's privacy policy.

Kipra Scout: AI Attack Surface Detector handles the following:

Web history

This developer declares that your data is

  • Not being sold to third parties, outside of the approved use cases
  • Not being used or transferred for purposes that are unrelated to the item's core functionality
  • Not being used or transferred to determine creditworthiness or for lending purposes
Search
Clear search
Close search
Google apps
Main menu