LibreChat vulnerabilities

LibreChat up to version 0.8.3 has a vulnerability in the MCP server integration that resolves ${VAR} placeholders from process.env during Zod schema validation of user-supplied MCP URLs. An authenticated user can create a malicious MCP configuration that sends critical secrets (CREDS_KEY, CREDS_IV, JWT_SECRET, MONGO_URI) to an attacker-controlled server.

LibreChat prior to v0.8.2-rc2 has a vulnerability in the MCP stdio transport that accepts arbitrary commands without validation, allowing any authenticated user to execute shell commands as root inside the container through a single API request. This vulnerability is fixed in v0.8.2-rc2.

LibreChat version 0.8.1-rc2 is prone to SSRF due to missing restrictions in the Actions feature by default. Agents can access internal services like the RAG API in the default Docker Compose setup.

A vulnerability in LibreChat prior to 0.8.4-rc1 allows an attacker with a stolen session token to regenerate all 2FA backup codes without requiring any TOTP token or existing backup code. This enables bypassing 2FA login or disabling 2FA entirely.

A vulnerability in LibreChat prior to 0.8.4-rc1 allows an authenticated user to bypass rate limiters added for the /fork endpoint by using the /duplicate endpoint, which performs the same expensive database operations but lacks any rate limiter. This can lead to server resource exhaustion.

LibreChat before version 0.8.4-rc1 allows authenticated users to set a custom URL for OpenAI-compatible API endpoints. This URL is used to construct HTTP requests without any SSRF validation, enabling traffic to be directed to internal network addresses.

A vulnerability in LibreChat before version 0.8.5 allows a malicious MCP server to steal access tokens intended for a legitimate server. The MCP OAuth implementation does not validate that the resource parameter from OAuth Protected Resource metadata (RFC 9728) matches the configured MCP server URL.

A vulnerability in LibreChat before version 0.8.4-rc1 allows any authenticated user to delete any other user's messages. The delete endpoint lacks a user constraint, enabling an attacker to use their own conversation ID and the victim's message ID to permanently remove messages.

LibreChat prior to version 0.8.4-rc1 has a vulnerability due to missing HTML escaping of double-quote characters in image alt text in Markdown. An attacker can inject malicious HTML code that executes in the victim's browser.

A vulnerability in LibreChat before version 0.8.4-rc1 allows an authenticated user to upload arbitrarily large files via the conversation import endpoint, potentially exhausting server disk space and memory. The issue stems from missing file size limits in a separate multer instance for this endpoint and a disabled application-level size check by default.

LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. In versions up to and including 0.8.3, a shared-agent editor can delete file records through `DELETE /api/files` that the owner has reused across multiple agents. The deletion removes the file globally — not just from the shared agent — breaking the owner's other private agents that reference the same `file_id`. The private agent retains a stale `file_id` reference that no longer resolves. A shared-agent editor can destroy files that the owner uses across multiple agents. The owner's private agents — which the attacker has no access to — break silently with stale `file_id` references. This is a cross-agent integrity violation: editing access to one agent should not affect another. Version 0.8.4 contains a patch.

LibreChat versions up to and including 0.8.3 allow users with only VIEW access to an MCP server to retrieve the server's decrypted admin-managed secrets through GET /api/mcp/servers and GET /api/mcp/servers/:serverName. The returned config includes plaintext values for apiKey.key and oauth.client_secret. This allows viewers of a shared MCP server to exfiltrate the underlying provider credentials. Version 0.8.4 contains a patch.

LibreChat is a ChatGPT clone supporting multiple AI providers. In versions up to 0.7.6, an IDOR vulnerability exists in the API keys management endpoint (PUT /api/keys). Due to the use of the spread operator after setting the authenticated user's ID, any authenticated user can inject a userId parameter to overwrite another user's API keys (e.g., OpenAI, Anthropic, Azure). An attacker can route victim's conversations through attacker-controlled keys or deny service.

LibreChat prior to version 0.8.4 trusts the 'name' field returned by the execute_code sandbox when persisting code-generated artifacts. An attacker can use traversal sequences (e.g., '../../') in the filename, leading to arbitrary file write on the server without validation.

LibreChat version 0.8.1-rc2 does not enforce proper access control when querying agent permissions. An authenticated attacker can read the permissions of arbitrary agents, even if they have no permissions for this agent, including permissions individually assigned to other users.

LibreChat version 0.8.1-rc2 does not enforce proper access control for file uploads to an agents file context and file search. An authenticated attacker with access to the agent ID can change the behavior of arbitrary agents by uploading new files, even without permissions for that agent. Fixed in version 0.8.2-rc2.

LibreChat in versions 0.8.0 and below lacks a handler for JSON parsing errors. SyntaxError from express.json() includes user input in the error message, which gets reflected in responses, creating an XSS risk if Content-Type isn't strictly enforced.

LibreChat versions 0.8.0 and below insufficiently validate JSON request bodies sent to the PATCH endpoint for prompt groups (/api/prompts/groups/:groupId). The patchPromptGroup function passes req.body directly to updatePromptGroup() without filtering sensitive fields, allowing users to modify prompts in unintended ways.

LibreChat is a ChatGPT clone with additional features. In versions 0.8.0 and below, when a user posts a question, the iconURL parameter of the POST request can be modified by an attacker. The malicious code is then stored in the chat which can then be shared to other users. When sharing chats with a potentially malicious 'tracker', resources loaded can lead to loss of privacy for users who view the chat link that is sent to them. This issue is fixed in version 0.8.1.

LibreChat is a ChatGPT clone with additional features. Prior to version 0.8.1-rc2, LibreChat is vulnerable to Server-side Request Forgery (SSRF), by passing specially crafted OpenAPI specs to its "Actions" feature and making the LLM use those actions. It could be used by an authenticated user with access to this feature to access URLs only accessible to the LibreChat server (such as cloud metadata services, through which impersonation of the server might be possible). This issue has been patched in version 0.8.1-rc2.

LibreChat version 0.7.9 is vulnerable to a Denial of Service (DoS) attack due to unbounded parameter values in the `/api/memories` endpoint. The `key` and `value` parameters accept arbitrarily large inputs without proper validation, leading to a null pointer error in the Rust-based backend when excessively large values are submitted. This results in the inability to create new memories, impacting the stability of the service.

In danny-avila/librechat version 0.7.9, there is an insecure API design issue in the 2-Factor Authentication (2FA) flow. The system allows users to disable 2FA without requiring a valid OTP or backup code, bypassing the intended verification process. This vulnerability occurs because the backend does not properly validate the OTP or backup code when the API endpoint '/api/auth/2fa/disable' is directly accessed.

A vulnerability in LibreChat version 0.7.9 allows HTML injection via the Accept-Language header. A logged-in user can send a crafted GET request, leading to arbitrary HTML injection into the <html lang=""> tag of the response, potentially enabling XSS attacks.

A mass assignment vulnerability exists in danny-avila/librechat, affecting all versions. This vulnerability allows attackers to manipulate sensitive fields by automatically binding user-provided data to internal object properties or database fields without proper filtering. As a result, any extra fields in the request body are included in agentData and passed to the database layer, allowing overwriting of any field in the schema, such as author, access_level, isCollaborative, and projectIds. Additionally, the Object.Prototype can be polluted due to the use of Object.assign with spread operators.

LibreChat contains an authorization bypass vulnerability due to improper access control checks. The checkAccess function uses permissions.some(), granting access when only one of multiple required permissions is present — a USER role can create agents despite lacking CREATE permission.