The Model Context Protocol Python SDK is the official Python SDK for Model Context Protocol servers and clients. From 1.9.1 until 1.30.0 and 2.2.0, OAuth client support in mcp.client.auth can send credentials to an authorization server chosen by an untrusted MCP server. ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider cannot identify the authorization server that owns their credentials. Affected HTTP clients using those providers, OAuthClientProvider, or the deprecated 1.x RFC7523OAuthClientProvider can disclose a client_secret, authorization code, PKCE code_verifier, or signed client assertion. MCP servers, stdio clients, and clients that attach their own tokens or headers are not affected. Interactive OAuthClientProvider use requires a person to start sign-in, while the unattended providers do not. For ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider, callers must configure issuer= after upgrading because omission still follows the MCP server's advertised authorization server, and the deprecated RFC7523OAuthClientProvider must be replaced because it has no issuer option. A stored client registration without an issuer remains unbound after upgrading, so stored OAuth client information must be cleared once to register again or a manually pre-registered record must be updated with its issuer. This issue is fixed in versions 1.30.0 and 2.2.0.
RedAmon before 7.8.0 contains a missing authentication vulnerability in the kali-sandbox terminal server that allows unauthenticated attackers to obtain a root shell by connecting to its WebSocket listener on port 8016. Attackers who can reach the port from co-resident containers or the host loopback can send keystrokes to an interactive /bin/bash session, executing commands as root inside the kali-sandbox container.
CVE-2026-125672 is a vulnerability in cockpit-storaged where a Tang keyserver address read from a clevis token in a LUKS header is concatenated into curl's argument vector without a proper '--' separator. This allows an attacker who can attach a crafted LUKS/clevis block device to cause curl to interpret an address starting with '-' as an option, leading to curl option injection. Exploitation requires a local unprivileged user with console or desktop login access and an administratively-privileged Cockpit user to interact with the affected device's Tang keyserver settings. The flaw can lead to arbitrary file read or write as the Cockpit session user. No complete mitigation exists without a code fix, but restricting local interactive access and verifying keyserver addresses can partially mitigate risk.
A vulnerability in dnf5daemon-server allows a local attacker to modify another user's live package-management session by exploiting missing verification of the session creator on D-Bus session objects. This can lead to the root daemon applying attacker-chosen package operations if the legitimate user authorizes execution. Red Hat Enterprise Linux is not affected as it does not ship dnf5daemon-server. Until a fix is available, users should avoid leaving interactive dnf5daemon sessions open on multi-user systems with untrusted local accounts and close sessions promptly after use.
CVE-2026-54162 affects Ember, a real-time terminal dashboard for Caddy servers with FrankenPHP support. Versions prior to 1.4.2 improperly handle terminal escape and control sequences in log fields, allowing unauthenticated remote clients to inject ANSI, OSC, or CSI control bytes via HTTP request URIs. When these logs are viewed in the interactive terminal UI, the terminal interprets these sequences, enabling spoofing of the monitoring dashboard, clipboard hijacking, and window-title spoofing. The issue is fixed in version 1.4.2.
Unrestricted Upload of File with Dangerous Type vulnerability in PT Norther Lights Production MapSVG mapsvg-lite-interactive-vector-maps allows Upload a Web Shell to a Web Server.This issue affects MapSVG: from n/a through 8.16.1.
A vulnerability in GIMP on Red Hat Enterprise Linux 10 involves an integer overflow when processing Bohemia Interactive (PAA) texture files. This flaw can cause the application to allocate an undersized memory buffer, leading to out-of-bounds reads and termination of the plug-in process. Exploitation requires a local user to open or preview a maliciously crafted file, limiting impact primarily to a denial of service or minor memory disclosure within the plug-in. GIMP is not installed by default on standard RHEL server environments, and the affected image format is uncommon, reducing exposure.
Mattermost versions 11.7.0, 11.8.0, 11.9.0, and 11.10.0 contain a vulnerability where action integration secrets are not properly stripped from post payloads before broadcasting post-deletion websocket events and returning thread API responses. This flaw allows authenticated channel members to access sensitive server-side webhook URLs and authentication context from interactive-message posts via the post_deleted websocket event or the threads REST API.
Microsoft and Health-ISAC disrupted a Chinese-language phishing marketplace comprising YAM Store, CoGUI, and Iron Man, which sent over one billion malicious messages since January 2026. These services supplied stolen accounts, phishing kits, and infrastructure to criminals worldwide, targeting consumers and more than 75,000 organizations in over 150 countries. The phishing kits allowed real-time credential harvesting with interactive platforms and included thousands of templates tailored to many countries. The disruption was enabled by a US federal court order and involved multiple cybersecurity and law enforcement partners. Microsoft disabled over 20,000 accounts and blocked more than 100,000 domains related to this operation. The phishing infrastructure depended on vendor licensing servers, making it vulnerable to takedown efforts.
A stack-based buffer overflow vulnerability exists in the interactive shell mode of the xmllint command-line tool used for parsing XML files in Red Hat AI Inference Server 3.5.2 (ROCm). This flaw occurs when an overly long command is input without proper size checks, potentially causing the program to crash. Exploitation requires local access and manual interaction with the tool, making it unlikely in typical environments. The impact is limited to availability due to potential crashes, with no known remote exploitation or code execution in common configurations.
The xrp-ledger-utils RubyGem version 0.9.1 contains malicious code embedded in its native extension builder. This code decodes and executes a base64-encoded Ruby payload that establishes a persistent remote shell connection to a hardcoded command-and-control server. The payload activates only on developer workstations, bypassing analysis environments, and provides an interactive shell to the attacker. The package itself has no legitimate functionality and is a malicious dropper disguised under unrelated project files.
CVE-2026-108258 is a path traversal vulnerability in py-shiny, a Python framework for interactive web applications. Versions from 1.4.0 up to but not including 1.6.4 allow an unauthenticated attacker to manipulate the bookmark restore feature by supplying a crafted state_id. This can lead to unauthorized reading of files outside the intended bookmark directory, including sensitive files like input.json and values.json. Additionally, in configurations using server-side bookmark storage and ui.input_file(), attackers can copy and expose arbitrary files from the server. The issue is fixed in version 1.6.4.
Heap-based buffer overflow in the legacy Blowfish encryption routine (BlowFishEncryptor::Encode, called by EncryptToString) in Progressive Robot hMailServer 6.0.0 through 6.3.5 allows an authenticated mailbox user to cause a denial of service (service crash), and possibly other unspecified impact. In 6.3.4 and 6.3.5, where the self-service REST API is enabled (it is off by default), the user does this remotely by adding a fetch account whose password is 129 to 247 characters long and not a multiple of 8, and then requesting their personal data export (GET /api/v1/me/export.zip), which encrypts that password with the legacy scheme. The same flaw is reachable on Windows by any local interactive user with no hMailServer credentials, through the COM method Utilities.BlowfishEncrypt, which checked no authentication. It is also reachable by every stored-secret write when ProtectStoredSecretsWithDPAPI is set to 0. For such a length, the routine's padding loop writes up to 7 zero bytes 2 to 232 bytes past the end of its 255-byte heap buffer. The ciphertext it returns is still correct.
Missing authorization on COM objects in Progressive Robot hMailServer 6.0.0 through 6.3.5 (Windows only) lets a local interactive user with no hMailServer credential read and write arbitrary files as the service account and queue mail as any sender. The service registers its COM classes with no DCOM access or launch permission and calls CoInitializeSecurity with no security descriptor, so any user logged on at the console or over Remote Desktop can activate the classes in the running service; a hMailServer.Message, its Attachments and Attachment, and a hMailServer.FetchAccount created this way carry a credential that never authenticated. Attachments.Add(path) and Attachment.SaveAs(path) performed no authorization check, and Message.Save/Copy and FetchAccount.AccountID/Save performed none either up to 6.3.3 and from 6.3.4 treated a holder with no credential as the server's own event-script host. Because the service does not impersonate the COM caller, Attachments.Add reads any file the service account can read and returns it, Attachment.SaveAs writes attacker-chosen bytes to any path it can write (on a LocalSystem installation, code execution as SYSTEM), Message.Save queues outbound mail from any address past the SMTP checks, and FetchAccount attaches a mail-fetch job to any mailbox. The objects an Application handed out behave the same once a later Authenticate on that Application fails.
### Summary
In affected versions, the SDK's OAuth client let the MCP server decide which authorization server received the client's OAuth credentials. Credentials were not tied to the authorization server they belong to.
A malicious or compromised MCP server could name its own authorization server. With no user interaction, the client would send it:
- the `refresh_token` and `client_secret` stored from an earlier sign-in
- the `client_secret` or signed assertion configured on a bundled provider
### Am I affected?
Yes, if both of these hold:
- your application uses the SDK's OAuth client over HTTP, through any of:
- an `authProvider` on a transport
- the `withOAuth()` middleware
- direct calls to `auth()` or `fetchToken()`
- it may connect to an MCP server you do not fully trust while holding credentials for a legitimate authorization server
Affected versions:
- `@modelcontextprotocol/sdk` 1.12.0 through 1.30.1
- `@modelcontextprotocol/client` 2.0.0 through 2.1.0, only for:
- bundled providers without `expectedIssuer`
- credentials stored or supplied without `issuer`
- direct calls to `fetchToken()`
- providers that read storage back through `OAuthTokensSchema` or `OAuthClientInformationSchema`
Not affected:
- MCP servers built with the SDK
- stdio clients
### Fix
Upgrade to:
- 1.x: `@modelcontextprotocol/sdk` 1.31.0 or later
- 2.x: `@modelcontextprotocol/client` 2.2.0 or later, and `@modelcontextprotocol/core` 2.2.0 or later if you import it directly
The client now records the authorization server as `issuer` on saved credentials and does not send them to a different one. The user signs in again, or the call throws.
In the following cases, upgrading to the patched version is not enough. You also need to make a change:
- **Bundled providers** (`ClientCredentialsProvider`, `PrivateKeyJwtProvider`, `StaticPrivateKeyJwtProvider`, `CrossAppAccessProvider`): pass `expectedIssuer`, for example `expectedIssuer: 'https://auth.example.com'`. Without it they still use whichever authorization server the MCP server names.
- **Credentials saved without `issuer`:** tokens and client information your `OAuthClientProvider` persisted (file, keychain, database) before upgrading, including everything 1.x saved before 1.31.0. They still go to whichever authorization server is named at first use. Add `issuer` to them, or clear them so users sign in again.
- **Your own `OAuthClientProvider`:** save exactly what `saveTokens()` and `saveClientInformation()` are given, including `issuer`. For pre-registered credentials, include `issuer` in what `clientInformation()` returns.
Not covered by this fix:
- a new interactive sign-in, which still goes to the authorization server the MCP server names. Only complete sign-ins for servers you trust.
- `refreshAuthorization()` and `exchangeAuthorization()` called directly
- 2.x with `skipIssuerMetadataValidation: true`
If an affected client may have connected to an untrusted MCP server, rotate its client secret or signing key and revoke its tokens.
If you cannot upgrade yet, connect OAuth clients only to MCP servers you trust. 2.0.0 and 2.1.0 already accept `expectedIssuer`.
CVE-2026-77185 is a critical authentication bypass vulnerability in Apache MINA SSHD versions 2.0.0 to 2.19.0 and 3.0.0-M1 to 3.0.0-M5. It affects the asynchronous authentication mechanism in the sshd-core component, potentially allowing signature checks to be skipped in public-key or hostbased authentication. The flaw arises from a logic error in the implementation of asynchronous authentication, which must be explicitly used by the server. Fixed in versions 2.20.0 and 3.0.0-M6, these updates also restrict asynchronous authentication to password or keyboard-interactive methods only, closing sessions if misused.