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
- Chainlit has command injection via MCP stdio transport that allows unauthenticated remote code execution The MCP stdio allowlist checks only the executable name, so an allowed launcher like npx can run arbitrary shell commands through its arguments.
- PraisonAI has a Browser Server WebSocket origin validation bypass via unanchored regex (patch bypass of CVE-2026-40289 / GHSA-8x8f-54wf-vv92) An unanchored regex lets forged WebSocket Origin headers pass the extension check, so the attacker can drive browser automation sessions.
- PraisonAI: Origin-validation bypass (startswith prefix match) enables unauthenticated cross-site request forgery against the PraisonAI MCP HTTP server The MCP HTTP server validates Origin with a prefix match, so a look-alike hostname can make a victim's browser send unauthenticated tool calls that write persistent rule files.
- PraisonAI has an origin validation bypass in MCP HTTP Stream transport that allows browser-mediated unauthenticated tool execution on local MCP server A prefix-based Origin check accepts attacker domains like localhost.evil.example, letting a malicious webpage call MCP tools on a local server that has no API key.
- PraisonAI MCP HTTP server has unauthenticated unbounded session accumulation (memory exhaustion; session TTL never enforced) The MCP HTTP server creates a session on every unauthenticated initialise request and never expires them, so clients can exhaust server memory.
- praisonaiagents has a `web_crawl` SSRF protection bypass via unchecked redirect targets The crawler checks only the first URL, then follows redirects unchecked, letting a public address send requests to internal or metadata endpoints.
- praisonaiagents vulnerable to SSRF in web_crawl tool via redirect-following and DNS rebinding (validate-then-fetch gap) The crawl tool checks only the first hostname resolution, then the HTTP client follows redirects and re-resolves DNS, letting requests reach internal addresses.
- chrome-devtools-mcp: validatePath() does not canonicalize symlinks before enforcing roots The workspace check compares unresolved paths, so an in-root symlink pointing outside the workspace passes and is then read or overwritten.
- Flowise: CVE-2025-8943 Patch Bypass: npm_config_yes bypasses MCP environment variable blocklist (Unauthenticated RCE) Setting npm_config_yes through the MCP server environment makes npx auto-install and run an arbitrary package, because only four environment names are blocked.
- MCP Ruby SDK: Unbounded session retention in StreamableHTTPTransport allows memory exhaustion via initialize flood Every initialize request creates a session kept in memory indefinitely unless the client deletes it, so repeated unauthenticated initializes exhaust server memory.
- MCP Ruby SDK: Ruby SSE Session Poisoning The Ruby MCP transport accepts tool calls for any session ID without checking ownership, so a stolen ID runs tools as the victim.
- MCP Ruby SDK: Unbounded line buffer in stdio transports leads to memory exhaustion (DoS) The MCP stdio transports read lines without a length limit, so a peer that sends bytes without a newline makes memory grow until the process is killed.
- MCP Ruby SDK: Unbounded JSON-RPC request body causes uncontrolled memory allocation in StreamableHTTPTransport The MCP Ruby HTTP transport reads and parses an unauthenticated POST body with no size limit, so one large request exhausts server memory.
- MCP Ruby SDK: Streamable HTTP transport lacks DNS-rebinding (Host/Origin) protection A webpage rebinds its hostname to localhost so the browser can call an unprotected local MCP Streamable HTTP server and read the results of its tools.
- GitHub MCP Server has Nil Pointer Dereference DoS in completion/complete Handler A completion/complete request missing its ref field makes the handler dereference nil, crashing the server for any unauthenticated client.
- LiteLLM: MCP Authentication Bypass via OAuth2 Passthrough Fallback An unauthenticated caller sends a fabricated Bearer token to the MCP endpoint, and a fallback path substitutes an empty auth object so the session is accepted.
- n8n: "Allowed HTTP Request Domains" Restriction Bypass via AI Agents MCP Connector The agent's MCP tool path ignores the allowed-domain restriction on shared credentials, so a member can send a secret to an attacker-chosen URL.
- n8n: Member-Level Users Can Execute Other Users' MCP Server Trigger Workflows via Missing OAuth Authorization Check A member-level user registers an OAuth client and approves consent for another user's MCP trigger workflow, obtaining a token that runs it with the owner's credentials.
- n8n: SSRF Protection Bypass via MCP Client Node The MCP Client node sends requests to user-supplied endpoints without applying SSRF checks or pinning resolved addresses, letting workflow users reach internal hosts.
- n8n: AI Agents Project Viewer Privilege Escalation via run_node_tool A read-only project user could chat with an agent whose node tool ran with project credentials, skipping execution and credential permission checks.
- MCP Python SDK: Experimental task handlers allow any client to access and cancel other clients' tasks Experimental MCP task handlers never checked which session created a task, so any connected client could list, read and cancel other clients' tasks.
- MCP Python SDK: HTTP transports serve session requests without verifying the authenticated principal The session lookup checks only the session ID, so a different authenticated client can inject messages into someone else's session.
- MCP Python SDK: WebSocket server transport does not support Host/Origin validation The WebSocket server transport accepts handshakes without checking Host or Origin headers, letting cross-origin pages connect to the MCP server.
- crewai-tools SSRF redirect bypass exposes internal services A URL that redirects or rebinds its DNS after the single validation lookup can reach internal hosts, because the check runs once and the original address is used.
- MCP Atlassian: DNS-rebinding TOCTOU bypass of the SSRF fix (CVE-2026-27826) The MCP server validates a header-supplied host's DNS result but later re-resolves it at connect time, so a rebinding name reaches internal addresses.
- mcp-atlassian: Arbitrary file read via missing path validation in confluence_upload_attachment The upload tool opens any caller-supplied path without the safety check its download counterpart uses, letting callers push host files to Confluence.
- mcp-atlassian: Arbitrary server-side file read via attachment upload A remote MCP client passes a server-side file path to the attachment upload tool, which reads that file from the server and uploads its contents.
- mcp-server-kubernetes argument injection can expose Kubernetes cluster credentials Leading-dash arguments evade the dangerous-flag check in kubectl tools, letting an attacker redirect kubectl to their server and receive the bearer token.
- GitHub MCP Server: Lockdown mode singleton in HTTP server causes cross-user GraphQL client confusion A process-wide cache keeps the first user's API client, so lockdown checks for later HTTP users run with another user's credentials.
- LiteLLM: MCP Proxy Has Improper Authentication A flaw in the MCP proxy's API key authentication lets a remote caller pass the check without valid credentials.
- BerriAI litellm has SSRF via Unvalidated spec_path URL in MCP OpenAPI Spec Loader An attacker-supplied spec path makes the MCP OpenAPI loader fetch arbitrary URLs from the server without restricting the destination.
- PraisonAI SandlockSandbox falls back to unrestricted subprocess execution when Landlock is unavailable When Landlock is unavailable, the sandbox silently runs submitted code through an unrestricted subprocess, ignoring file and network limits.
- PraisonAI: Missing Authentication for Critical Function and Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection') in praisonai An unauthenticated UI endpoint accepts a command and arguments for an MCP stdio connection, so the server launches attacker-chosen local processes.
- PraisonAI ToolsMCPServer legacy SSE transport accepts attacker Host/Origin and exposes registered tools The legacy SSE MCP server exposes its tools without checking Host, Origin, or authentication, so a DNS-rebinding page can list and invoke them.
- PraisonAI: SpiderTools redirect-target SSRF protection bypass Redirects are followed without validating the target, so a public URL can redirect the scraper to internal or metadata addresses.
- PraisonAI: MCP SSE transport binds 0.0.0.0 with no authentication and no Origin validation; bundled SecurityConfig is never wired in The MCP SSE server listens on all interfaces with no authentication or origin checks, so any reachable host can call every registered tool.
- Chrome DevTools for agents: daemon.pid write follows symlinks in /tmp fallback runtime directory The MCP daemon writes its PID file without O_NOFOLLOW, so a local user can plant a symlink in the fallback temp path and truncate a victim-writable file.
- n8n: MCP Browser HTTP Transport Exposes Unauthenticated Browser-Control Sessions The HTTP transport of an MCP browser server accepts sessions and tool calls without authentication, exposing browser control to network clients.
- Claude Code GitHub Action Secret Exposure
- PraisonAI vulnerable to unauthenticated arbitrary file read via MCP workflow.show, workflow.validate, deploy.validate An unauthenticated MCP tools/call to workflow.show passes a filename straight to file reads, since the containment fix covered other handlers but not this one.
- MCP Server Kubernetes: Tool Access Control Bypass via Presentation-Layer Filtering Without Execution-Layer Enforcement The MCP server filters tools only in the listing step, so a client that knows a tool name can call it directly, ignoring the restriction.
- Open WebUI's chat completion API allows tool restrictions to be bypassed The chat completion endpoint runs any server tool named by ID without checking the caller's permission, using the server's stored credentials.
- Flowise has an MCP Security Bypass that Enables RCE A custom MCP server config running docker build with a remote URL slips past a flag blocklist that omits build, so a remote Dockerfile executes commands.
- PraisonAI MCP `tools/call` path-traversal => RCE via Python `.pth` injection Path-traversing rule names in MCP tool calls write a .pth file into site-packages, so attacker Python code runs on every later Python start.
- PraisonAI has unauthenticated RCE via `tool_override.py` (CVE-2026-40287 patch bypass) The recipe runner dynamically executes a tools.py found beside a recipe template, and this loader path lacks the opt-in gate added to other tool loaders.
- rmcp Streamable HTTP server transport has a DNS rebinding vulnerability A browser page can reach a locally running Streamable HTTP MCP server through DNS rebinding because the Host header is never checked, enabling tool calls.
- n8n Vulnerable to Unauthenticated Denial of Service via MCP Client Registration The MCP OAuth client registration endpoint accepts unauthenticated large payloads and stores them without limits, exhausting server memory even when MCP is disabled.
- n8n Vulnerable to XSS via MCP OAuth client An MCP OAuth client name containing script content is rendered unescaped in a notification, running in another user's session when they see it.
- n8n has Open Redirect in MCP OAuth Consent Flow Unauthenticated OAuth client registration lets an attacker set an arbitrary redirect URI, which the consent deny handler follows without validation.
- LiteLLM: Authenticated command execution via MCP stdio test endpoints MCP stdio test endpoints spawn attacker-supplied commands and only require a valid key, not the admin role, so any user can run commands.