mcp-remote is exposed to OS command injection when connecting to untrusted MCP servers due to crafted input from the authorization_endpoint response URL

Vulnerability in dirhistory plugin Description: the widgets that go back and forward in the directory history, triggered by pressing Alt-Left and Alt-Right, use functions that unsafely execute eval on directory names. If you cd into a directory with a carefully-crafted name, then press Alt-Left, the system is subject to command injection. Impacted areas: - Functions pop_past and pop_future in dirhistory plugin.

Cursor is a code editor built for programming with AI. In versions and below, a vulnerability in the Cursor CLI Beta allowed an attacker to achieve remote code execution through the MCP (Model Context Protocol) server mechanism by uploading a malicious MCP configuration in .cursor/mcp.json file in a GitHub repository. Once a victim clones the project and opens it using Cursor CLI, the command to run the malicious MCP server is immediately executed without any warning, leading to potential code execution as soon as the command runs. This issue is fixed in version 2025.09.17-25b418f.

Cursor is a code editor built for programming with AI. In versions 1.7.28 and below, an input validation flaw in Cursor's MCP server installation enables specially crafted deep-links to bypass the standard security warnings and conceal executed commands from users if they choose to accept the server. If an attacker is able to convince a victim to navigate to a malicious deeplink, the victim will not see the correct speedbump modal, and if they choose to accept, will execute commands specified by the attackers deeplink.

Atril is a simple multi-page document viewer. Atril is vulnerable to a critical Command Injection Vulnerability. This vulnerability gives the attacker immediate access to the target system when the target user opens a crafted document or clicks on a crafted link/URL using a maliciously crafted CBT document which is a TAR archive. A patch is available at commit ce41df6.

PaddlePaddle before 2.6.0 has a command injection in get_online_pass_interval. This resulted in the ability to execute arbitrary commands on the operating system.

PaddlePaddle before 2.6.0 has a command injection in convert_shape_compare. This resulted in the ability to execute arbitrary commands on the operating system.

PaddlePaddle before 2.6.0 has a command injection in _wget_download. This resulted in the ability to execute arbitrary commands on the operating system.

A flaw was found in SSSD, where the sssctl command was vulnerable to shell command injection via the logs-fetch and cache-expire subcommands. This flaw allows an attacker to trick the root user into running a specially crafted sssctl command, such as via sudo, to gain root access. The highest threat from this vulnerability is to confidentiality, integrity, as well as system availability.

OS Command Injection vulnerability in bbultman gitsome through 0.2.3 allows attackers to execute arbitrary commands via a crafted tag name of the target git repository.

Poddycast is a podcast app made with Electron. Prior to version 0.8.1, an attacker can create a podcast or episode with malicious characters and execute commands on the client machine. The application does not clean the HTML characters of the podcast information obtained from the Feed, which allows the injection of HTML and JS code (cross-site scripting). Being an application made in electron, cross-site scripting can be scaled to remote code execution, making it possible to execute commands on the machine where the application is running. The vulnerability is patched in Poddycast version 0.8.1.

PaddlePaddle before 2.5.0 has a command injection in fs.py. This resulted in the ability to execute arbitrary commands on the operating system.

lifion-verify-dependencies through 1.1.0 is vulnerable to OS command injection via a crafted dependency name on the scanned project's package.json file.

Activity Watch is a free and open-source automated time tracker. Versions prior to 0.11.0 allow an attacker to execute arbitrary commands on any macOS machine with ActivityWatch running. The attacker can exploit this vulnerability by having the user visiting a website with the page title set to a malicious string. An attacker could use another application to accomplish the same, but the web browser is the most likely attack vector. This issue is patched in version 0.11.0. As a workaround, users can run the latest version of aw-watcher-window from source, or manually patch the `printAppTitle.scpt` file.

git-commiters is a Node.js function module providing committers stats for their git repository. Prior to version 0.1.2, there is a command injection vulnerability in git-commiters. This vulnerability manifests with the library's primary exported API: gitCommiters(options, callback) which allows specifying options such as cwd for current working directory and revisionRange as a revision pointer, such as HEAD. However, the library does not sanitize for user input or practice secure process execution API to separate commands from their arguments and as such, uncontrolled user input is concatenated into command execution. This issue has been patched in version 0.1.2.

A flaw was found with the libssh API function ssh_scp_new() in versions before 0.9.3 and before 0.8.8. When the libssh SCP client connects to a server, the scp command, which includes a user-provided path, is executed on the server-side. In case the library is used in a way where users can influence the third parameter of the function, it would become possible for an attacker to inject arbitrary commands, leading to a compromise of the remote target.

In Zmanda Management Console 3.3.9, ZMC_Admin_Advanced?form=adminTasks&action=Apply&command= allows CSRF, as demonstrated by command injection with shell metacharacters. This may depend on weak default credentials.

Adobe After Effects version 18.1 (and earlier) is affected by a potential Command injection vulnerability when chained with a development and debugging tool for JavaScript scripts. An unauthenticated attacker could leverage this vulnerability to achieve arbitrary code execution in the context of the current user. Exploitation of this issue requires user interaction in that a victim must open a malicious file.

GNU Bash through 4.3 bash43-026 does not properly parse function definitions in the values of environment variables, which allows remote attackers to execute arbitrary commands via a crafted environment, as demonstrated by vectors involving the ForceCommand feature in OpenSSH sshd, the mod_cgi and mod_cgid modules in the Apache HTTP Server, scripts executed by unspecified DHCP clients, and other situations in which setting the environment occurs across a privilege boundary from Bash execution. NOTE: this vulnerability exists because of an incomplete fix for CVE-2014-6271, CVE-2014-7169, and CVE-2014-6277.

An issue was discovered in Centreon before 18.10.8, 19.10.1, and 19.04.2. It allows CSRF with resultant remote command execution via shell metacharacters in a POST to centreon-autodiscovery-server/views/scan/ajax/call.php in the Autodiscovery plugin.

The "gitDiff" function in Wayfair git-parse <=1.0.4 has a command injection vulnerability. Clients of the git-parse library are unlikely to be aware of this, so they might unwittingly write code that contains a vulnerability. The issue has been resolved in version 1.0.5.

A vulnerability exists that could allow the execution of operating system commands on systems running MiPlatform 2019.05.16 and earlier. An attacker could execute arbitrary remote command by sending parameters to WinExec function in ExtCommandApi.dll module of MiPlatform.

This affects the package @graphql-tools/git-loader before 6.2.6. The use of exec and execSync in packages/loaders/git/src/load-git.ts allows arbitrary command injection.

Fortra BoKS Manager contains an OS command injection vulnerability in the client upgrade and patch tooling for legacy tar-based client installations. A malicious or compromised legacy tar-installed client selected for upgrade or patching may be able to cause commands to be executed on the BoKS Master during client version handling.

Improper neutralization of special elements used in an OS command ('OS command injection') vulnerability in TUBITAK BILGEM Software Technologies Research Institute Pardus OS My Computer allows OS Command Injection. This issue affects Pardus OS My Computer: from <=0.7.5 before 0.8.0.

radare2-mcp version 1.6.0 and earlier contains an os command injection vulnerability that allows remote attackers to execute arbitrary commands by bypassing the command filter through shell metacharacters in user-controlled input passed to r2_cmd_str(). Attackers can inject shell metacharacters through the jsonrpc interface parameters to achieve remote code execution on the host running radare2-mcp without requiring authentication.

Cherry Studio is a desktop client that supports for multiple LLM providers. In version 1.5.1, a remote code execution (RCE) vulnerability exists in the Cherry Studio platform when connecting to streamableHttp MCP servers. The issue arises from the server’s implicit trust in the oauth auth redirection endpoints and failure to properly sanitize the URL. This issue has been patched in version 1.5.2.

Agentic-Flow is an AI agent orchestration platform. Prior to 2.0.14, agentic-flow MCP server tools in src/mcp/standalone-stdio.ts, src/mcp/fastmcp/servers/claude-flow-sdk.ts, src/mcp/fastmcp/servers/stdio-full.ts, src/mcp/fastmcp/servers/http-streaming-updated.ts, src/mcp/fastmcp/servers/http-sse.ts, src/mcp/fastmcp/servers/poc-stdio.ts, src/mcp/fastmcp/tools/agent/{execute,list,parallel}.ts, src/mcp/fastmcp/tools/swarm/orchestrate.ts, and src/mcp/fastmcp/tools/hooks/pretrain.ts interpolated attacker-influenceable tool parameters such as agent, task, name, language, and agentdb directly into shell command strings passed to execSync(), allowing arbitrary OS command execution with the privileges of the MCP server user. This issue is fixed in version 2.0.14.

Nest is a framework for building scalable Node.js server-side applications. In versions 0.2.0 and below, a critical Remote Code Execution (RCE) vulnerability was discovered in the @nestjs/devtools-integration package. When enabled, the package exposes a local development HTTP server with an API endpoint that uses an unsafe JavaScript sandbox (safe-eval-like implementation). Due to improper sandboxing and missing cross-origin protections, any malicious website visited by a developer can execute arbitrary code on their local machine. The package adds HTTP endpoints to a locally running NestJS development server. One of these endpoints, /inspector/graph/interact, accepts JSON input containing a code field and executes the provided code in a Node.js vm.runInNewContext sandbox. This is fixed in version 0.2.1.

gpsd through release-3.27.5, fixed at commit 4c06658, contains a command injection vulnerability in gpsprof that allows attackers who control the GPS device subtype value to execute arbitrary shell commands by embedding backtick payloads in the gnuplot plot title without proper escaping. The subtype field sourced from a DEVICES JSON log entry or NMEA PGRMT sentence is written into a generated gnuplot program via a set title statement with only double-quote characters escaped, enabling arbitrary shell command execution as the user running gnuplot when the victim renders the generated plot through the gpsprof and gnuplot workflow.

Claude Code is an agentic coding tool. From 2.1.38 until 2.1.163, Claude Code's worktree handling allowed creation of worktrees named ".git" and navigation to worktrees outside the sandbox context, enabling git directory confusion attacks. By exploiting symlink manipulation and git fsmonitor execution during worktree operations, an attacker could overwrite files in the user's home directory (such as .zshenv), leading to code execution outside of seatbelt sandbox restrictions. Reliably exploiting this required the user to clone a malicious repository containing prompt injection content and run Claude Code against it. This vulnerability is fixed in 2.1.163.

The Zoom Client before 4.4.53932.0709 on macOS allows remote code execution, a different vulnerability than CVE-2019-13450. If the ZoomOpener daemon (aka the hidden web server) is running, but the Zoom Client is not installed or can't be opened, an attacker can remotely execute code with a maliciously crafted launch URL. NOTE: ZoomOpener is removed by the Apple Malware Removal Tool (MRT) if this tool is enabled and has the 2019-07-10 MRTConfigData.

Markdown Preview Enhanced before 0.8.28 opens external files and links from the preview through a shell and does not validate untrusted inputs taken from the markdown document - the diagram filename attribute, imported file paths, and the latex_engine code-chunk attribute. On Windows, a crafted markdown document can inject operating system commands that execute when the document is previewed. Fixed in 0.8.28 by passing these inputs as literal arguments instead of through a shell and validating them before use.

Ghidra versions prior to 12.0.3 improperly process annotation directives embedded in automatically extracted binary data, resulting in arbitrary command execution when an analyst interacts with the UI. Specifically, the @execute annotation (which is intended for trusted, user-authored comments) is also parsed in comments generated during auto-analysis (such as CFStrings in Mach-O binaries). This allows a crafted binary to present seemingly benign clickable text which, when clicked, executes attacker-controlled commands on the analyst’s machine.

Cursor is a code editor built for programming with AI. In versions 1.17 through 1.2, there is a UI information disclosure vulnerability in Cursor's MCP (Model Context Protocol) deeplink handler, allowing attackers to execute 2-click arbitrary system commands through social engineering attacks. When users click malicious `cursor://anysphere.cursor-deeplink/mcp/install` links, the installation dialog does not show the arguments being passed to the command being run. If a user clicks a malicious deeplink, then examines the installation dialog and clicks through, the full command including the arguments will be executed on the machine. This is fixed in version 1.3.

Yocto Project is an open source collaboration project that helps developers create custom Linux-based systems regardless of the hardware architecture. In Yocto Projects Bitbake before 2.6.2 (before and included Yocto Project 4.3.1), with the Toaster server (included in bitbake) running, missing input validation allows an attacker to perform a remote code execution in the server's shell via a crafted HTTP request. Authentication is not necessary. Toaster server execution has to be specifically run and is not the default for Bitbake command line builds, it is only used for the Toaster web based user interface to Bitbake. The fix has been backported to the bitbake included with Yocto Project 5.0, 3.1.31, 4.0.16, and 4.3.2.

phpFileManager 0.9.8 allows remote attackers to execute arbitrary commands via a crafted URL.

BentoML is a Python library for building online serving systems optimized for AI apps and model inference. Prior to 1.4.39, a malicious bentofile.yaml containing a newline-injected value in envs[*].name produces unquoted RUN directives in the BentoML-generated Dockerfile. When the victim runs bentoml containerize on the imported bento, those RUN directives execute on the host during docker build. This vulnerability is fixed in 1.4.39.

PraisonAI is a multi-agent teams system. Prior to 4.5.121, the execute_command function and workflow shell execution are exposed to user-controlled input via agent workflows, YAML definitions, and LLM-generated tool calls, allowing attackers to inject arbitrary shell commands through shell metacharacters. This vulnerability is fixed in 4.5.121.

PraisonAIAgents is a multi-agent teams system. Prior to 1.5.128, he memory hooks executor in praisonaiagents passes a user-controlled command string directly to subprocess.run() with shell=True at src/praisonai-agents/praisonaiagents/memory/hooks.py. No sanitization is performed and shell metacharacters are interpreted by /bin/sh before the intended command executes. Two independent attack surfaces exist. The first is via pre_run_command and post_run_command hook event types registered through the hooks configuration. The second and more severe surface is the .praisonai/hooks.json lifecycle configuration, where hooks registered for events such as BEFORE_TOOL and AFTER_TOOL fire automatically during agent operation. An agent that gains file-write access through prompt injection can overwrite .praisonai/hooks.json and have its payload execute silently at every subsequent lifecycle event without further user interaction. This vulnerability is fixed in 1.5.128.

An undocumented debug CGI endpoint in T3 Technology CPE models T625Pro v1.0.07, T6825G v1.0.03 allows unauthenticated attackers to execute arbitrary system commands as root via supplying a crafted HTTP query string.

@pensar/apex <= 0.0.58 is vulnerable to OS command injection via the smart_enumerate tool. The createSmartEnumerateTool() function in src/core/agent/tools.ts constructs a shell command by concatenating unsanitized values from the extensions array and url parameter into a string passed to Node.js child_process.exec(). Because exec() spawns a shell, shell metacharacters in those values are interpreted by the host shell, resulting in arbitrary OS command execution with the privileges of the running process. NOTE: this is disputed by the Supplier because the report is about intended behavior, as explained in the Security Policy of the pensarai/apex GitHub repo.

The shell tool command allowlist in the SecurityPolicy of OpenHuman desktop agent through 0.54.0 (default Supervised security policy) can be bypassed to execute arbitrary OS commands with the privileges of the desktop user. Two flaws in src/openhuman/security/policy.rs combine: (1) is_args_safe() blocks the find flags -exec and -ok but not the functionally identical -execdir and -okdir, which also execute an arbitrary command for each matched file; and (2) skip_env_assignments() strips leading inline KEY=value environment-variable assignments before allowlist validation, so a command such as GIT_EXTERNAL_DIFF=<cmd> git diff is validated as the allowed git diff but, when executed via the shell, runs <cmd> through git's environment-driven hooks (for example GIT_EXTERNAL_DIFF or GIT_SSH_COMMAND). Because the sandbox is the primary trust boundary between untrusted LLM-processed content and the host operating system, an attacker can achieve remote code execution via indirect prompt injection: a malicious document, email, calendar event, or web page ingested by the agent instructs it to run a benign-looking allowlisted command, resulting in arbitrary command execution, data exfiltration, arbitrary file read/write, and lateral movement on the user's machine. The issue was fixed in commit 60050aa09a870f53ed7e4cd40ed41fd2860329e7 (first released in 0.54.22-staging; first stable release 0.56.0), which blocks -execdir/-okdir for find.

Cursor is a code editor built for programming with AI. Prior to 2.0 ,if a visited website contains maliciously crafted instructions, the model may attempt to follow them in order to “assist” the user. When combined with a bypass of the command whitelist mechanism, such indirect prompt injections could result in commands being executed automatically, without the user’s explicit intent, thereby posing a significant security risk. This vulnerability is fixed in 2.0.

In lookatme (python/pypi package) versions prior to 2.3.0, the package automatically loaded the built-in "terminal" and "file_loader" extensions. Users that use lookatme to render untrusted markdown may have malicious shell commands automatically run on their system. This is fixed in version 2.3.0. As a workaround, the `lookatme/contrib/terminal.py` and `lookatme/contrib/file_loader.py` files may be manually deleted. Additionally, it is always recommended to be aware of what is being rendered with lookatme.

aws-mcp-server MCP server is vulnerable to command injection. An attacker can craft a prompt that once accessed by the MCP client will run arbitrary commands on the host system.

OpenClaw is a personal AI assistant. Versions 2026.1.8 through 2026.2.13 have a command injection in the maintainer/dev script `scripts/update-clawtributors.ts`. The issue affects contributors/maintainers (or CI) who run `bun scripts/update-clawtributors.ts` in a source checkout that contains a malicious commit author email (e.g. crafted `@users[.]noreply[.]github[.]com` values). Normal CLI usage is not affected (`npm i -g openclaw`): this script is not part of the shipped CLI and is not executed during routine operation. The script derived a GitHub login from `git log` author metadata and interpolated it into a shell command (via `execSync`). A malicious commit record could inject shell metacharacters and execute arbitrary commands when the script is run. Version 2026.2.14 contains a patch.

Versions of the package degit before 2.8.6, from 3.0.0 and before 3.3.1 are vulnerable to Command Injection due to improper sanitisation of user input for git shell commands directly invoked with exec() method by _cloneWithGit() and fetchRefs() functions. An attacker can execute arbitrary operating system commands as the process user by supplying a specially crafted git repository name.

Pi-Hole 4.3 allows Command Injection.