MCP Atlassian versions prior to 0.22.0 contain a vulnerability in the Jira and Confluence attachment upload tools that allows reading arbitrary local files on the server. This occurs because these tools accept arbitrary filesystem paths and send the selected file contents to Atlassian, enabling attackers in HTTP or multi-user environments to access sensitive files such as configuration or credential data. The issue is fixed in version 0.22.0.
A vulnerability in MCP Atlassian prior to version 0.22.0 allows arbitrary server-local file uploads to Jira and Confluence attachments via unrestricted file_path parameters. This flaw enables permitted clients in remote or multi-user deployments to disclose host files without requiring shell or direct filesystem access. The issue is fixed in version 0.22.0.
CVE-2026-89039 is a path traversal vulnerability in Grafana's mcp-k6 component that allows an authenticated caller with limited privileges to read arbitrary files on the server. By invoking the convert_playwright_script prompt and passing a crafted file path, an attacker can bypass directory restrictions using symbolic links and access sensitive files such as SSH keys and cloud credentials. This vulnerability affects versions 0.3.0 through before 0.7.0 of mcp-k6.
The rmcp OAuth client in the modelcontextprotocol rust-sdk accepts a server-controlled resource_metadata URL from the WWW-Authenticate header and fetches it without validating origin or blocking internal network addresses. This allows an attacker-controlled or compromised MCP server to induce the client to make HTTP requests to internal or cloud metadata endpoints, potentially exposing sensitive internal resources. The vulnerability is classified as SSRF with medium severity, depending on the network context of the embedding application. A patch is available to enforce same-origin checks and block private network addresses.
CVE-2026-104120 is a server-side request forgery (SSRF) vulnerability in the fetch_url function of the mcp_server_fetch/server.py file within the modelcontextprotocol mcp-server-fetch and mcp-server-everything components up to version 2026.6.4. This vulnerability allows remote attackers to manipulate URL or path arguments to induce the server to make unintended requests. The vulnerability has been publicly disclosed, and a fix is pending acceptance of a pull request.
CVE-2025-71427 is a path traversal vulnerability in GongRzhe Office-PowerPoint-MCP-Server versions up to 2.0.7. It allows attackers to read and write files outside the intended working directory by using absolute paths or directory traversal sequences. This can be exploited by steering an AI agent via prompt injection to abuse certain functions, potentially overwriting server-writable files or loading external files.
CVE-2026-97662 is an argument injection vulnerability in the diff scan operation of AWS security-agent-mcp-server versions 0.1.1 up to but not including 0.2.0. This flaw allows context-dependent threat actors to create, overwrite, or truncate arbitrary files on the host system outside the intended workspace by supplying a crafted reference value. The issue is fixed in version 0.2.0.
The AI Engine – The Chatbot, AI Framework & MCP for WordPress plugin for WordPress is vulnerable to Stored Cross-Site Scripting in versions up to, and including, 3.8.0 This is due to a chain of missing input neutralization and output escaping across the /mwai-ui/v1/chats/submit REST endpoint, the PHP error-log parser (MeowKit_MWAI_Helpers::php_error_logs), the Advisor task (Meow_MWAI_Modules_Advisor::run_advisor), and the Advisor dashboard widget (advisor_metabox): the server-parameter denylist in chat_submit strips only exact key names such as 'model' while convert_keys() later canonicalizes 'model_' back to 'model', allowing an unauthenticated caller to place an attacker-controlled string (including CR/LF) into $query->model; final_checks() throws an Exception whose message embeds that raw string, and the non-streaming, non-admin catch branch writes it to the PHP error log unmodified — creating a forged log line that the plugin's own parser subsequently returns as recent PHP-error content; run_advisor() then appends that content verbatim to the AI prompt (indirect prompt injection — CWE-1427), the returned JSON is stored in the mwai_advisor_data option with no schema validation or HTML sanitization, and advisor_metabox() concatenates the resulting 'title' and 'description' values directly into the WordPress dashboard widget without esc_html(), wp_kses(), or equivalent escaping. This makes it possible for unauthenticated attackers to inject arbitrary web scripts in pages that will execute whenever an administrator accesses the WordPress dashboard.
CVE-2026-102906 is an OS command injection vulnerability in the 0xshariq github-mcp-server affecting the child_process.exec function in src/github.ts. The vulnerability allows remote attackers to execute arbitrary OS commands by manipulating the File argument. The project uses a rolling release model, so specific affected or fixed versions are not available. The issue was reported early but has not yet received a response from the project. Public exploit code exists, but no known exploitation in the wild has been confirmed.
CVE-2026-102878 is a high-severity vulnerability in mcp-chrome-bridge up to version 1.0.31. It involves an origin validation error in the native-server HTTP API that allows attackers to bypass Cross-Origin Resource Sharing (CORS) restrictions. This flaw enables malicious web pages to make unauthorized cross-origin requests to the local server, potentially invoking browser automation features such as script execution, reading page content, and capturing screenshots.
mark3labs mcp-filesystem-server version 0.11.1 contains a directory traversal vulnerability due to improper handling of dangling symbolic links in the validatePath function. This flaw allows certain file operations to create or modify files outside the intended allowed directories by following unresolved symlinks. The vulnerability has a medium severity with a CVSS score of 5.9.
PyJWT versions from 2.13.0 up to but not including 2.14.0 contain a vulnerability where deeply nested JSON arrays in the JWT header cause a RecursionError during parsing. This error is not caught by the library's usual exception handling, resulting in an HTTP 500 server error and a full traceback logged. This allows unauthenticated attackers to cause a denial of service on JWT validation endpoints without crashing the server or exposing data. The issue is fixed in PyJWT 2.14.0.
CVE-2026-55156 is a path traversal vulnerability in the token-optimizer-mcp product by ooples. Versions prior to 5.1.0 have an unauthenticated HTTP server exposing endpoints that accept a sessionId parameter concatenated directly into a filesystem path. This allows an attacker to read arbitrary .jsonl files on the server filesystem. The issue is fixed in version 5.1.0.
CVE-2026-55157 is a high-severity OS command injection vulnerability in the ooples token-optimizer-mcp product prior to version 5.1.0. The flaw exists in the smart_user tool, where the username argument of the get-user-info operation can be exploited to execute arbitrary shell commands with the privileges of the token-optimizer-mcp server user. This vulnerability has been patched in version 5.1.0.
fast-mcp-telegram versions prior to 30.1 contain a server-side request forgery (SSRF) vulnerability in the send_message and send_message_to_phone MCP tools. These tools accept files as lists of HTTP(S) URLs, which the server downloads and attaches to outgoing Telegram messages. The URL validation relies on a denylist that checks only the hostname string without DNS resolution, allowing attackers to bypass restrictions and cause the server to fetch resources from internal or private network addresses. This results in a full-read SSRF that can exfiltrate data by returning the fetched content as a Telegram file attachment. The vulnerability is patched in version 30.1.
Obot before v0.23.0 contains a server-side request forgery vulnerability in remote MCP server registration that allows privileged users to specify arbitrary URLs without destination validation. Attackers with Power User or higher roles can coerce Obot to make requests to internal services and cloud metadata endpoints, reading responses in error messages to disclose sensitive credentials.
Obot versions before v0.23.0 fail to enforce authentication on MCP Registry endpoints under /v0.1/* when registry authentication is enabled. Unauthenticated attackers can read registry metadata including server names, descriptions, repository URLs, and connect URLs by sending GET requests to /v0.1/servers.