Mar 1, 2026·6 min read·430 visits
CVE-2025-66478 (React2Shell) is a critical (CVSS 10.0) unauthenticated Remote Code Execution vulnerability in Next.js applications using the App Router. By manipulating the React Flight protocol serialization stream, attackers can invoke the JavaScript `Function` constructor to execute arbitrary code. Active exploitation has been observed. Immediate patching and secret rotation are required.
A critical vulnerability, designated CVE-2025-66478 (and tracking the upstream React CVE-2025-55182), exists in the React Server Components (RSC) implementation within Next.js. The flaw resides in the 'Flight' protocol, used for serializing component trees between the server and client. Specifically, the server-side deserializer fails to properly validate reference types in `multipart/form-data` payloads, allowing attackers to access the global `Function` constructor via prototype pollution gadgets. This leads to arbitrary code execution on the server without authentication.
CVE-2025-66478 represents the Next.js-specific tracking identifier for a critical deserialization flaw in the React Server Components (RSC) architecture, widely referred to as 'React2Shell'. The vulnerability stems from the upstream React library (tracked as CVE-2025-55182) but primarily impacts Next.js applications due to the framework's default exposure of Server Actions and the App Router.
The core issue lies within the React Flight Protocol, a streaming protocol used to transport the Virtual DOM and component data from the server to the client. When a Next.js application processes a request (typically a POST request to a Server Action), it utilizes a streaming decoder to parse the incoming payload. This decoder supports a reference mechanism allowing parts of the stream to refer to previously defined chunks.
The vulnerability is a classic Deserialization of Untrusted Data (CWE-502). The decoder allows property access on referenced objects without sufficient validation. An attacker can craft a malicious payload that abuses this reference system to traverse the prototype chain (__proto__, constructor) and access the JavaScript Function constructor. Once the constructor is accessible, the attacker can instantiate it with arbitrary code strings and execute them within the Node.js context of the server.
The vulnerability is rooted in the resolveModel function within the react-server-dom-webpack and react-server-dom-turbopack packages. The React Flight protocol allows payloads to contain references, denoted by $ syntax (e.g., $1), which point to other chunks in the stream.
In the vulnerable implementation, the decoder permitted references to include property accessors, such as $1:key. While intended for legitimate object traversal, the logic failed to block access to sensitive internal properties. An attacker can supply a reference formatted as $1:constructor:constructor. In JavaScript, the constructor of a basic object is Object, and the constructor of Object is Function. This provides a direct handle to the global Function constructor, which is functionally equivalent to eval() when called.
Furthermore, the protocol supports a 'Bind' operator, denoted as $B, intended for lazy loading and binding arguments to functions. By chaining the polluted Function constructor reference with the $B handler, an attacker can force the server to invoke the constructed function immediately upon deserialization. The combination of unchecked property access during reference resolution and the ability to execute bound functions creates a reliable gadget chain for Remote Code Execution.
The attack leverages the specific structure of multipart/form-data requests that Next.js accepts for Server Actions. The attack flow proceeds as follows:
$@0 to refer to the first form field.$1:constructor:constructor. The decoder resolves this by taking the object at index 1 and walking up its prototype chain to the Function constructor.$B (bind) to bind the malicious string (e.g., require('child_process').execSync(...)) to the Function constructor and executes it.The lack of an allowlist for property accessors in the protocol parser is the definitive root cause.
Exploitation requires no authentication and can be performed against any Next.js application using the App Router (default in newer versions) that exposes a Server Action endpoint. Even if no explicit Server Actions are defined, the internal router endpoints may still be accessible.
Payload Structure:
The exploit uses a specifically crafted JSON payload passed via multipart/form-data. The critical component is the _formData reference abuse.
POST / HTTP/1.1
Host: target.com
Next-Action: arbitrary_id
Content-Type: multipart/form-data; boundary=----WebKitFormBoundary7MA4YWxkTrZu0gW
------WebKitFormBoundary7MA4YWxkTrZu0gW
Content-Disposition: form-data; name="0"
{"then":"$1:__proto__:then","status":"resolved_model","value":"{\"then\":\"$B1337\"}","_chunks":"$Q2","_formData":{"get":"$1:constructor:constructor"}}
------WebKitFormBoundary7MA4YWxkTrZu0gW
Content-Disposition: form-data; name="1"
"$@0"
------WebKitFormBoundary7MA4YWxkTrZu0gW--Mechanism:
In this payload, the _formData property exploits the get handler to resolve $1:constructor:constructor. The payload is designed to satisfy the thenable checks within React's internal promise handling, forcing the resolution of the malicious function.
Indicators of Compromise:
Successful exploitation results in the command specified in the payload (e.g., whoami, curl) being executed with the privileges of the Node.js process running the Next.js server. This often leads to immediate reverse shell connections or exfiltration of environment variables.
The impact of CVE-2025-66478 is rated Critical (CVSS 10.0) because it grants full control over the server application without any prior credentials.
Confidentiality (High): Attackers can read all environment variables (process.env), which typically contain database connection strings, API keys (AWS, Stripe, OpenAI), and signing secrets. This allows lateral movement to other infrastructure components.
Integrity (High): Attackers can modify application source code, inject persistent backdoors (e.g., adding a middleware that logs all user credentials), or modify data in connected databases.
Availability (High): The attacker can crash the server, delete critical files, or deploy ransomware.
Since Next.js applications are often exposed directly to the internet (unlike internal microservices), the attack surface is vast. The vulnerability affects the default configuration of the widely adopted App Router.
Remediation requires upgrading the next package to a patched version. Vercel and the React team have released fixes that strictly validate references in the Flight protocol, blocking access to constructor and prototype.
Patched Versions:
Immediate Actions:
npm install next@latest or yarn add next@latest immediately.npx fix-react2shell-next.Temporary WAF Mitigation:
If patching is not immediately possible, configure your WAF to block POST requests containing Next-Action headers where the body contains strings like $B, $@, or constructor:constructor. However, WAF rules are easily bypassed by obfuscation; patching is the only complete solution.
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H| Product | Affected Versions | Fixed Version |
|---|---|---|
Next.js Vercel | 15.0.0 - 15.0.4 | 15.0.5 |
Next.js Vercel | 15.1.0 - 15.1.8 | 15.1.9 |
Next.js Vercel | 16.0.0 - 16.0.6 | 16.0.7 |
| Attribute | Detail |
|---|---|
| CWE ID | CWE-502 |
| Attack Vector | Network (HTTP POST) |
| CVSS | 10.0 (Critical) |
| Exploit Status | Active / Weaponized |
| Protocol | React Flight (RSC) |
| Impact | Remote Code Execution (RCE) |
The application deserializes untrusted data without sufficiently verifying that the resulting data will be valid.
An authentication bypass and account takeover vulnerability in the AshAuthentication Elixir library (developed by team-alembic) allows unauthenticated remote attackers to compromise local accounts. By relying on mutable and unverified email claims instead of stable cryptographic issuer and subject pairings during OAuth2 and OIDC federated login flows, the application fails to validate the trust boundary of the incoming session.
A critical logical flaw in the eml_parser Python module prior to version 3.0.2 allows malicious URLs to evade automated security analysis pipelines. By encoding key URI delimiter characters as HTML decimal entities, an attacker can mask indicators of compromise. Security controls, orchestration layers, and sandbox systems fail to detect these links, while downstream Mail User Agents natively reconstruct the malicious hyper-references when processed by end-users. This mechanism undermines the integrity of automated indicator extraction processes within Security Operations Centers.
A denial of service vulnerability in GOVCERT-LU eml_parser before version 3.0.2 allows unauthenticated remote attackers to trigger an unhandled RecursionError exception. The issue arises during the parsing of structured email headers containing excessively nested parentheses representing Comments and Folding White Space (CFWS). Because the parser fails to catch this recursion-limit exception from Python's standard library, processing of the entire mail immediately aborts, which can disrupt automated security triage pipelines and email ingestion components.
Prior to version 3.0.2, GOVCERT-LU's eml_parser library is vulnerable to an algorithmic complexity Denial of Service (DoS) vulnerability via the comment-stripping routine noparenthesis() in routing.py. An unauthenticated attacker can submit a crafted EML file containing nested parenthesized comments to cause complete CPU saturation. This happens due to a quadratic time complexity bottleneck in regex replacement of nested structures.
Whistle prior to version 2.10.3 contains a path traversal vulnerability in its internal service layer. An unauthenticated remote attacker can read arbitrary files on the hosting operating system by issuing a crafted GET request containing relative or absolute file paths to the `/cgi-bin/temp/get` endpoint. This behavior occurs because the application fails open when an input file parameter does not match the temporary file format regex.
An arbitrary file read and write vulnerability exists in the Model Context Protocol (MCP) server endpoints of sublinear-time-solver and consciousness-explorer. By providing unvalidated file paths to the export_state, import_state, saveVectorToFile, and loadVectorFromFile tools, local attackers can read or overwrite sensitive host files.