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
- Model Context Protocol Python SDK: HTTP client transports followed cross-origin redirects carrying custom headers and bodies The MCP HTTP client follows redirects to other origins, forwarding custom credential headers and request bodies to hosts the user never configured.
- Model Context Protocol Python SDK: Client fetched server-chosen $ref URLs while validating tool results against output schemas A malicious MCP server embeds remote or file references in a tool's output schema, making the client's validator fetch them on every call.
- Model Context Protocol Python SDK: Server bearer-token authentication accepts tokens issued for a different resource server The bearer token verifier does not check the audience, so tokens issued for other services are accepted by the MCP server.
- MCP TypeScript SDK: HTTP server transports and handlers did not limit request body size or JSON-RPC batch length MCP HTTP servers accept unbounded POST bodies and JSON-RPC batches, so oversized requests exhaust memory and block other clients.
- FastMCP: Post-transform component authorization is skipped on direct calls A server transform changes component metadata after authorisation, and direct lookups return the transformed component without rechecking policy, so denied tools and resources can be reached.
- FastMCP MCP client wrappers construct M2M OAuth providers without issuer binding — client secret redirectable to attacker-advertised authorization server An untrusted MCP server can advertise a different authorization server, so the client sends its OAuth client secret or signed assertion to the attacker.
- MCP TypeScript SDK: Experimental tasks were not tied to the session that created them The MCP TypeScript SDK's experimental task store did not tie tasks to their creating session, letting other clients list, read and cancel them.
- MCP TypeScript SDK: One HTTP request within the default size limit could exhaust server memory during validation The MCP TypeScript SDK validates unlimited array elements and object keys, so one HTTP request within the size limit exhausts server memory.
- MCP TypeScript client SDK: HTTP client transports and OAuth requests followed cross-origin redirects and re-sent custom headers and request bodies Cross-origin redirects cause the client to resend custom headers and request bodies, including OAuth secrets, to an attacker-chosen origin.
- MCP TypeScript SDK: Server bearer authentication accepted tokens issued for another service The MCP TypeScript SDK's bearer auth never checked token audience, so tokens issued for other services were accepted by the MCP server.
- MCP Python SDK: Unbounded Session Accumulation in StreamableHTTP Transport Enables Resource Exhaustion Streamable HTTP sessions are never expired or capped, so a client opening many short sessions grows server memory without limit.
- PraisonAI: MCP HTTP-stream transport is unauthenticated by default, exposing tool enumeration and an unvalidated tool-call surface The MCP HTTP server skips authentication when no key is configured and forwards tool-call arguments without checking them against the advertised schema.
- PraisonAI: Plugin Auto-Discovery Executes Arbitrary Python Files Without Verification The plugin manager automatically imports and runs any Python file found in project or home plugin directories without verification.
- Langflow: OS command injection (RCE) via arbitrary command in MCP stdio server configuration Langflow launches arbitrary commands from MCP stdio server configurations without validation, so anyone who can add a server runs OS commands on the host.
- Langflow: IP Spoofing Bypass via `X-Forwarded-For` Allowing Remote Configuration Write A spoofed X-Forwarded-For header makes the MCP install endpoint treat a remote request as local, allowing an MCP client config file to be overwritten.
- Langflow has Authenticated Cross-Project File Disclosure via Unscoped MCP Resource Handlers An authenticated user sends a crafted resource URI to their own project's MCP endpoint and reads another user's flow file, because the read is not authorised.
- PraisonAI: Crawl4AI/Chromium backend is also affected by the `web_crawl` SSRF validation bypass The crawl4ai browser backend re-resolves DNS and follows redirects without the per-connection SSRF guard, letting requests reach internal addresses.
- MCP TypeScript SDK: OAuth client could send credentials to an authorization server chosen by the MCP server A malicious MCP server can choose its own authorization server, so the client automatically sends stored refresh tokens and client secrets to it.
- rmcp OAuth client fetches server-controlled resource_metadata URLs An MCP OAuth client fetches a resource metadata URL taken from a server's WWW-Authenticate header without checking whether it points to internal addresses.
- MCP Atlassian: Arbitrary File Read & Exfiltration (Confused Deputy) in JIRA update_issue An MCP tool passes agent-supplied absolute paths straight to an upload call, so the privileged server reads restricted host files and attaches them to a ticket.
- MCP Atlassian: ENABLED_TOOLS / Toolset authorization bypass Tool filters are applied only when listing tools, so a direct tools/call request can invoke any registered tool by name, bypassing the configured restriction.
- MCP Atlassian: Insecure File Permissions on OAuth Token Storage The MCP server writes OAuth tokens to a file with default permissions, so other local users can read the Atlassian credentials.
- MCP Atlassian: SSRF via DNS Rebinding in Header-Based Authentication Flow The header-based auth flow validates a Jira URL's DNS result, then the later request resolves again, so a rebinding host reaches internal addresses.
- mcp-atlassian has an incomplete SSRF remediation The MCP middleware accepts Jira and Confluence base URLs from request headers without validation, so the server fetches attacker-chosen internal hosts.
- MCP Atlassian: SSRF redirect protection missing for basic-auth and OAuth authentication branches The SSRF redirect check is attached only to the header token auth branch, so basic-auth and OAuth requests follow redirects to internal addresses such as cloud metadata.
- MCP Atlassian: Arbitrary file read via confluence_upload_attachment allows exfiltration of server credentials The Confluence upload tool opens any caller-supplied local path without checks, letting a client attach server files such as credentials to a page.
- MCP Atlassian: Incomplete path traversal fix allows intra-CWD module overwrite and RCE (bypass of GHSA-xjgw-4wvw-rgm4) An incomplete traversal fix still lets attachment downloads overwrite application source modules in the working directory, so injected code runs on restart.
- MCP Atlassian: Arbitrary file read/exfiltration via upload_attachment missing validate_safe_path() The MCP attachment upload tool sends local files to Confluence without the path check its download methods use, so an agent can be steered to read any file.
- MCP Atlassian: Unauthenticated arbitrary local file read via upload_attachment file_path, chained with missing auth on streamable-http transport Omitting the Authorization header lets unauthenticated clients call MCP tools, and an upload tool then reads any server file path and exposes its contents.
- MCP Atlassian: Arbitrary File Read via Upload Attachment Tools The upload attachment tools accept arbitrary server file paths without the safe-path check, letting an MCP client or an assistant steered by injected content exfiltrate files.
- MCP Atlassian: Arbitrary local file READ via unconstrained file_path in upload_attachment (Confluence + Jira) The Confluence and Jira attachment upload tools open any local path supplied to them and upload its contents, with no base-directory restriction.
- MCP Atlassian: Arbitrary server-local file upload to Jira/Confluence attachments via unrestricted file_path parameters MCP attachment upload tools take caller-supplied local paths, read those files from the server, and upload them to Jira or Confluence without restriction.
- MCP Atlassian: Reflected XSS in OAuth Setup Callback Handler The local OAuth callback server of an MCP integration echoes the error query parameter into HTML unescaped, so a crafted link runs script in the browser.
- MCP Atlassian: OAuth fallback token storage writes plaintext access and refresh tokens with group-readable permissions A plaintext fallback token file is created with the default umask, leaving access and refresh tokens readable by other local users.
- MCP Atlassian: Path traversal in upload_attachment allows arbitrary file read (incomplete fix for CVE-2026-27825) The upload MCP tools open any caller-supplied file path without the confinement check that the download path received in an earlier fix.
- MCP Atlassian: SSRF Protection Bypass A backslash-based userinfo trick makes the URL validator check a harmless hostname while the actual request goes to a loopback or internal address.
- MCP Atlassian: Path traversal in upload_attachment allows arbitrary file read and exfiltration via MCP tool call An authenticated MCP client passes an arbitrary local path to upload_attachment, which reads the file and sends its contents to Atlassian, because the upload path skips the safe-path check.
- MCP Atlassian: HTTP upload tools accept arbitrary server-local file paths Caller-supplied server-local paths are read by the upload tool and attached to Jira or Confluence, so a prompt-injected agent can exfiltrate local secrets.
- MCP Atlassian: Path Traversal / Arbitrary File Read in confluence_upload_attachment MCP tool (incomplete fix of GHSA-xjgw-4wvw-rgm4) An MCP upload tool opens any local file path a caller supplies and posts its contents to an attacker-chosen Confluence host, since the earlier fix only covered downloads.
- MCP Atlassian: Incomplete fix for GHSA-7r34-79r5-rcc9: redirect-based SSRF via unhooked requests session in Jira user-permission lookup A public base URL passes the initial internal-address check, then a redirect to an internal host is followed by a request path that skips the redirect validation hook.
- MCP Atlassian: Jira and Confluence attachment upload tools can read arbitrary server-local files Upload tools accept arbitrary server-side paths and read those files without restriction, then send their contents to Atlassian as attachments.
- MCP Atlassian: JIRA_PROJECTS_FILTER / CONFLUENCE_SPACES_FILTER allow forbidden-project content exfiltration (one LIVE-proven on Atlassian Cloud) Substring matching on JQL and CQL project filters lets crafted queries, including those from prompt-injected agents, read content from projects outside the allowlist.
- [mcp-atlassian] Authentication bypass in HTTP transport: AtlassianOpaqueTokenVerifier accepts any non-empty token The HTTP transport's token verifier accepts any non-empty string, and the default setup applies no auth challenge, so unauthenticated requests reach Atlassian operations.
- MCP Atlassian: MCP HTTP Client Server-Local File Exfiltration via Unvalidated Attachment Upload Path An MCP attachment tool accepts any server-side file path, and attacker-supplied service headers redirect the upload so local files are sent to an attacker endpoint.
- RMCP: Custom HTTP headers leak to cross-origin redirect targets The Rust MCP client forwards custom auth headers on cross-origin 307 and 308 redirects, so an attacker-controlled target receives them.
- RMCP: Missing Resource Field Validation in OAuth Protected Resource Metadata Discovery The MCP client skips checking the resource field in protected resource metadata, so a malicious server can steer OAuth to another authorization server.
- RMCP: Unauthenticated permanent session-table leak in rmcp Streamable HTTP server transport leads to remote denial-of-service Unauthenticated non-initialise POSTs allocate sessions that are never released, so remote requests steadily exhaust server memory.
- Open WebUI: A user's session cookies are sent to tool servers configured for bearer authentication A tool call to a bearer-authenticated server can carry the user's session cookies because a shared callable reads a cookie jar built for a different connection.
- n8n: Agent Workflow Tool Bypasses Sub-Workflow Caller Policy Agent tool invocations skip the sub-workflow caller restriction, letting a user who builds an agent run workflows the owner had locked down and read their output.
- Chainlist has SSRF via MCP SSE and streamable-http transports that allows unauthenticated internal network access The unauthenticated MCP endpoint accepts arbitrary URLs and headers for SSE and streamable-HTTP transports, so the server sends requests to attacker-chosen internal targets.