Browse / MCP and Tool Server Flaws
MCP and Tool Server Flaws
Vulnerabilities in Model Context Protocol servers, clients, plugins and agent tool integrations. 177 records. Reference: OWASP Agentic · LLM03.
- Authentication and access control 38
- Code and command execution 37
- Path traversal and file access 27
- Server-side request forgery 23
- Denial of service 14
- Data exfiltration 10
- Information exposure 8
- Other weaknesses 6
- Cross-site scripting 5
- Cross-site request forgery 3
- Sandbox escape 2
- Credentials from AI Agent Configuration 1
- AI Agent Tool Poisoning: Implementation 1
- AI Agent Tool Poisoning: Definition and Instructions 1
- Discover AI Agent Configuration: Tool Definitions 1
- Claude Code: Trust Dialog Bypass via Git Worktree Spoofing Allows Arbitrary Code Execution A crafted repository's worktree commondir file points to an already trusted path, so the trust dialog is skipped and repo-defined hooks run immediately.
- Langflow vulnerable to injection The X-Forwarded-For header is used unsanitised when generating MCP configuration, so attacker-supplied values are injected into that output.
- PraisonAI has an incomplete fix for CVE-2026-34935 - OS Command Injection MCP server command parsing hands arbitrary executables and inline-code flags to subprocesses because the earlier fix added no allowlist or argument checks.
- Flowise: Authenticated RCE Via MCP Adapters The MCP stdio configuration accepts any command and only filters some patterns, so an allowed runner like npx can be given code-execution flags.
- OpenAI Codex CLI enables code execution through malicious MCP (Model Context Protocol) configuration files Running the Codex CLI in a malicious repository auto-loads project config that defines commands, which execute without user confirmation.
- MCP Server Kubernetes has an Argument Injection in port_forward tool via space-splitting The port-forward tool splits a space-joined kubectl command, so spaces inside user-supplied fields inject extra kubectl flags.
- PraisonAI Vulnerable to Implicit Execution of Arbitrary Code via Automatic `tools.py` Loading PraisonAI executes any tools.py found in the working directory at startup without consent, so planting that file runs attacker code.
- PraisonAI Browser Server allows unauthenticated WebSocket clients to hijack connected extension sessions An unauthenticated WebSocket client that omits the Origin header can register as a controller and drive automation sessions on other connected browser extensions.
- PraisonAI Vulnerable to RCE via Automatic tools.py Import The agent framework automatically imports a tools.py file from the working directory, so any planted file runs its code.
- PraisonAI Vulnerable to Sensitive Environment Variable Exposure via Untrusted MCP Subprocess Execution MCP stdio servers launched by the framework inherit the entire host environment, so untrusted server commands can read secrets such as API keys.
- PraisonAIAgents: Path Traversal via Unvalidated Glob Pattern in list_files Bypasses Workspace Boundary An agent file-listing tool validates the directory but not the glob pattern, so `..` segments reveal metadata for files outside the workspace.
- PraisonAI Has Path Traversal in FileTools The file tool checks for parent-directory sequences only after normalisation has removed them, so the check never fires and any file is reachable.
- DNS Rebinding Protection Disabled by Default in Model Context Protocol Go SDK for Servers Running on Localhost Localhost HTTP MCP servers lack DNS rebinding protection by default, letting a malicious website invoke tools through a rebound domain.
- Claude SDK for Python: Memory Tool Path Validation Race Condition Allows Sandbox Escape A symlink swapped in after path validation in the SDK's memory tool lets a local attacker redirect reads and writes outside the memory directory.
- PraisonAI Has ReDoS via Unvalidated User-Controlled Regex in MCPToolIndex.search_tools() Tool search compiles a caller-supplied string as a regex with no limits, so nested quantifiers backtrack catastrophically and block the thread.
- PraisonAI Has Authentication Bypass via OAuthManager.validate_token() validate_token accepts any bearer token when the token store is empty, so unauthenticated callers reach every MCP tool, including agent execution.
- PraisonAI: OS Command Injection in MCPHandler.parse_mcp_command() The PraisonAI CLI passes the --mcp argument to a process launcher without validation, so any supplied command runs as the user.
- PraisonAI Has SSRF in FileTools.download_file() via Unvalidated URL An agent file download tool passes an arbitrary URL to an HTTP client that follows redirects, so requests can reach internal metadata services and hosts.
- mcp-handler has a tool response leak across concurrent client sessions ('Race Condition') Reusing a stateless transport across concurrent clients in the MCP SDK leaks one client's auth context and tool results to another.
- FastMCP has a Command Injection vulnerability - Gemini CLI Shell metacharacters in an MCP server name get interpreted by cmd.exe when the install command runs a .cmd wrapper on Windows.
- FastMCP: Missing Consent Verification in OAuth Proxy Callback Facilitates Confused Deputy Vulnerabilities The FastMCP OAuth proxy callback does not verify fresh user consent, so a lured browser session yields an authorization code for the attacker's client.
- FastMCP OpenAPI Provider has an SSRF & Path Traversal Vulnerability Path parameters are inserted into URL templates without encoding, so ../ sequences escape the API prefix and the server sends authenticated requests to arbitrary backend URLs.
- MCP Ruby SDK: Insufficient Session Binding Allows SSE Stream Hijacking via Session ID Replay An SSE stream is bound only to a session ID, so anyone holding that ID can replace the stream and receive the victim's messages.
- Cross-Site Tool Execution for HTTP Servers without Authorizatrion in github.com/modelcontextprotocol/go-sdk The HTTP transport accepts cross-site POST requests without checking Origin or requiring a JSON content type, so a web page can trigger MCP tool calls.
- Improper handling of null Unicode character when parsing JSON in github.com/modelcontextprotocol/go-sdk Trailing null characters on JSON keys are treated as equal to base names, so duplicate-key messages can override fields and evade inspection by proxies.
- FastMCP OAuth Proxy token reuse across MCP servers The OAuth proxy issues tokens bound to its base URL rather than the requested resource, so a token minted for one MCP server can be replayed against another.
- MCP Atlassian has SSRF via unvalidated X-Atlassian-Jira-Url / X-Atlassian-Confluence-Url headers Unauthenticated requests carrying custom URL headers make the MCP server send requests to attacker-chosen URLs, including internal metadata services.
- MCP Atlassian has an arbitrary file write leading to arbitrary code execution via unconstrained download_path in confluence_download_attachment The attachment download tool writes content to any caller-chosen path without boundary checks, so an attacker can plant files such as cron entries that get executed.
- Azure MCP Server has Server-Side Request Forgery issue that allows authorized attacker to elevate privileges over a network An authorised user can make the Azure MCP Server send server-side requests to arbitrary destinations, gaining higher privileges over the network.
- MCP Go SDK Vulnerable to Improper Handling of Case Sensitivity The Go MCP SDK accepted case-variant JSON keys, so an intermediary matching exact field names could be evaded by a malicious peer's messages.
- mcp-server-git : Path traversal in git_add allows staging files outside repository boundaries The git_add tool accepts relative paths with parent-directory sequences and stages files outside the repository, which can later be committed and pushed.
- @modelcontextprotocol/sdk has cross-client data leak via shared server/transport instance reuse Reusing one transport across clients makes colliding JSON-RPC IDs route responses to the wrong client's stream.
- Claude Code has a Domain Validation Bypass which Allows Automatic Requests to Attacker-Controlled Domains The WebFetch trusted-domain check used prefix matching, so attacker domains that begin with a trusted name were fetched automatically, enabling data leakage.
- OpenCode's Unauthenticated HTTP Server Allows Arbitrary Command Execution An agent's local HTTP server has no authentication and permissive CORS, so any local process or web page can run shell commands.
- Microsoft Playwright MCP Server vulnerable to DNS Rebinding Attack; Allows Attackers Access to All Server Tools The MCP server does not validate the Origin header, so a web page using DNS rebinding can invoke its tools from the victim's browser.
- Anthropic's MCP TypeScript SDK has a ReDoS vulnerability Crafted URIs matched against exploded resource template patterns trigger catastrophic regex backtracking in the MCP SDK URI template matcher, hanging the server.
- FastMCP updated to MCP 1.23+ due to CVE-2025-66416 FastMCP pinned MCP SDK versions below 1.23, which contain a known vulnerability, so upgrading FastMCP to 2.14.0 or later fixes it.
- mcp-server-git's unrestricted git_init tool allows repository creation at arbitrary filesystem locations The git_init MCP tool created repositories at any path the server could reach without validation, making arbitrary directories eligible for later git operations.
- mcp-server-git argument injection in git_diff and git_checkout functions allows overwriting local files Tool arguments starting with a dash are passed straight to git, so a crafted value is read as an option that writes output to an attacker-chosen file.
- mcp-server-git has missing path validation when using --repository flag The git MCP server did not check repo path arguments in tool calls against the configured repository, so calls could operate on other repositories.
- mcp-server-kubernetes has potential security issue in exec_in_pod tool String commands given to exec_in_pod are run through a shell without validation, so metacharacters inject extra commands into the pod.
- Model Context Protocol (MCP) Python SDK does not enable DNS rebinding protection by default Localhost MCP HTTP servers lack DNS rebinding protection by default, letting a malicious website reach their tools through the browser.
- Model Context Protocol (MCP) TypeScript SDK does not enable DNS rebinding protection by default A malicious webpage uses DNS rebinding to reach an unauthenticated MCP server on localhost, because host validation is off by default.
- FastMCP Auth Integration Allows for Confused Deputy Account Takeover The MCP server issues authorization on behalf of an upstream identity provider without binding it to the real user, letting an attacker take over other accounts.
- FastMCP vulnerable to reflected XSS in client's callback page The OAuth callback page of the FastMCP client inserts the error query parameter unescaped, so a crafted link runs attacker script in the callback origin.
- FastMCP vulnerable to windows command injection in FastMCP Cursor installer via server_name A server name containing cmd metacharacters is placed into a Windows start command, so the installer runs attacker-chosen processes.
- Flowise is vulnerable to arbitrary file exposure through its ReadFileTool The agent's read_file tool passes a caller-supplied path straight to the file store, so an authenticated user can read any file on the host.
- Flowise is vulnerable to arbitrary file write through its WriteFileTool The agent's write-file tool accepts any destination path without restricting it to the working directory, letting an authenticated user write files anywhere on the server.
- Flowise vulnerable to RCE via Dynamic function constructor injection The Custom MCP node passes the server config it receives into a dynamic function constructor, so attacker-supplied config runs JavaScript on the host.
- figma-developer-mcp vulnerable to command injection in get_figma_data tool The MCP tool builds a shell command string from unvalidated parameters passed to child_process.exec, so shell metacharacters in tool inputs run arbitrary commands.