Is my AI tool safe?Protocols and SDKsAnthropic and the MCP community
Is Model Context Protocol safe?
No tool is safe or unsafe on its own: it depends on what it can reach and how it is set up. The record for Model Context Protocol: 20 incidents in this center's library and 12 CVEs in its advisory feed (3 critical, 9 high), the newest published Oct 15, 2026. Below: what went wrong, what was fixed, and how to use it safely.
What to watch for
The plumbing of the agent ecosystem, and the busiest category in the library: malicious servers in registries, tool descriptions that change after review, STDIO configs that are command execution, and an SDK that handed OAuth secrets to any server that asked.
How to use it safely
Install servers only from vetted sources, pin each one to its tool list and re-approve on change, treat server config as code, and keep the SDKs on current versions.
What it can reach
- Installs plugins, skills or tools from registries
- Runs commands with your permissions
- Holds API keys and tokens
- Reads attacker-controlled content
What went wrong: incidents in the library
- Corelight: a pinned MCP server can still pull fresh dependencies on every launch, so pinning the server does not pin what runs Medium
James Lagermann of Corelight published on November 10, 2026 that an MCP server configured for one fixed release, with its configuration untouched, ran with 15 different dependency sets over five weeks; 23 of its 69 packages… - Symantec honeypot: an attack agent hijacks exposed AI endpoints to plan intrusions and plant miners, at the owner's expense High
Symantec reported on October 21, 2026 that its AI honeypot, which poses as an unprotected internet-facing AI service, captured in summer 2026 the full instructions of an attack agent whose operators wanted a real model to drive… - INC and The Gentlemen affiliates: exposed servers show ransomware intrusions run through OpenCode, commercial models and 32 MCP servers High
SOCRadar's Threat Research Unit published on October 20, 2026, and updated on October 23, its analysis of seven exposed servers used by affiliates of the INC and The Gentlemen ransomware groups. It has moderate to high… - OpenCode on Windows: a remote MCP server's sign-in link could run PowerShell commands as the developer Critical
Eilon Cohen of Pillar Security disclosed a command injection (CWE-78) in OpenCode, the coding agent maintained by Anomaly, in versions 1.0.137 through 1.18.32 on Windows and WSL; macOS and Linux are not affected. When OpenCode… - OpenAI's official Render skill sent production apps' tokens to a domain that did not exist, until researchers registered it High
A skill in OpenAI's official GitHub repository of skills, which teaches agents how to deploy applications on Render, told production apps to send their tokens to render-internal.com, a domain nobody had registered, CSO Online… - Langflow: a logged-in user's MCP server command runs as a shell, and honeypots have counted 2,300 exploit attempts since September Critical
CVE-2026-105740 is an authenticated remote code execution flaw in Langflow, rated CVSS 9.9 in the GitHub advisory GHSA-7w94-79vh-5mr2 that Langflow published on September 10, 2026: any logged-in user can run operating system… - PoeLLM: a miner that found its command server in a poem infected 3,400 exposed LiteLLM, Ollama and developer servers High
Lumen's Black Lotus Labs reported on October 7, 2026 that malware it calls PoeLLM had compromised more than 3,400 servers since at least April, with up to 800 active in a single day and at least 11 command servers,… - MCP TypeScript SDK: a malicious MCP server could choose where the client's OAuth secrets were sent High
The official MCP TypeScript SDK's OAuth client did not tie credentials to the authorization server they belong to, so a malicious or compromised MCP server could name its own authorization server and the client would send it,… - Loom for AWS: an unauthenticated client could take over the agent control plane, and MCP scopes could leak OAuth secrets Critical
AWS published security bulletin 2026-124 on October 2, 2026 for Loom for AWS, an AWS Labs open-source platform for orchestrating AI agents. CVE-2026-103956 (CVSS 10.0): when no identity provider was configured, the… - MCP Python SDK: a malicious server could steal the client's OAuth credentials High
Cycode reported that in the official MCP Python SDK a malicious MCP server could redirect the client's OAuth token exchange to an endpoint it controlled and capture the client secret, the authorization code and the PKCE proof… - Obot: the Docker quickstart gave anyone on the network admin rights and the Docker socket Critical
Obot, an open-source AI agent and MCP platform, documented a Docker quickstart that listened on 0.0.0.0:8080 with authentication turned off, so every request got the Owner and Admin roles, and that mounted the host's Docker… - LiteLLM: any Bearer token opened the MCP gateway, and attackers used it High
LiteLLM is the open-source gateway many teams put in front of their model providers and MCP servers. Before version 1.84.0, its MCP Streamable HTTP endpoint accepted any Bearer token: when key validation failed, the handler fell… - Deadbugz: an MCP server that turns malicious after three tool calls High
Pillar Security found an active campaign submitting a fake "productivity-suite" MCP server to public MCP directories. It behaves normally at first; after a client makes three tool calls it changes the instructions it gives the… - GhostJacking: firewall logs become instructions for the agent that reads them High
At DEF CON 34, Tenet Security showed that a request a web application firewall blocks is logged with its payload intact, and an agent asked to investigate blocked traffic reads that log as instruction rather than evidence.… - MCP STDIO design flaw: server configs become command execution Critical
OX Security found that MCP's STDIO transport runs whatever command a server configuration names. Any AI app that lets users, web requests or an agent change that configuration can be turned into remote code execution, and the… - SmartLoader clones an Oura Ring MCP server to spread an infostealer High
Straiker found the SmartLoader malware operation cloning a legitimate Oura Ring MCP server, backing it with at least five fake GitHub accounts and forks to look popular, and listing the trojanized copy in MCP registries.… - postmark-mcp: first malicious MCP server found in the wild High
An npm package posing as Postmark's MCP server, which lets AI assistants send email, quietly blind-copied every message it sent to an attacker. Researchers called it the first real-world malicious MCP server. - MCPoison: a swapped MCP config gives persistent code execution in Cursor High
Cursor asked developers to approve an MCP server once and then trusted it forever, even if its command changed. A collaborator could commit a harmless config, wait for approval, then swap in a malicious command. - Supabase MCP: a support ticket tricks a Cursor agent into leaking secrets High
A support ticket with hidden instructions could make a developer's Cursor agent read a secret database table and post it back into the ticket for the attacker, because the agent connected with the all-powerful service_role key. - GitHub MCP: a malicious issue makes an agent leak private repositories High
Anyone who could open an issue on a public repository could trick a developer's AI agent into copying data from their private repositories into a public pull request. The flaw was in how agents combine tools and permissions, not…
Known CVEs
- CVE-2026-111768 High MCP TypeScript SDK runs out of memory validating one large tool request
- CVE-2026-104850 High MCP TypeScript SDK sends OAuth credentials to an authorization server the MCP server chooses
- CVE-2026-111767 High MCP TypeScript SDK experimental tasks not tied to the session that created them
- CVE-2026-25536 High MCP TypeScript SDK cross-client data leak on shared server instances
- CVE-2026-23744 Critical MCPJam inspector unauthenticated RCE through an exposed HTTP endpoint
- CVE-2025-66414 High MCP TypeScript SDK lacks DNS rebinding protection by default
- CVE-2025-66416 High MCP Python SDK lacks DNS rebinding protection by default
- CVE-2025-53967 High Framelink Figma MCP server command injection in get_figma_data
- CVE-2025-6514 Critical mcp-remote OS command injection from an untrusted MCP server
- CVE-2025-53355 High mcp-server-kubernetes command injection in several tools
- CVE-2025-53109 High MCP filesystem server reads unintended files via symlinks
- CVE-2025-49596 Critical MCP Inspector remote code execution through missing authentication
How many people use it
- 68.8M npm downloads of @modelcontextprotocol/sdk in a week (Sep 23 to 29, 2026) · npm download counts
- 219M PyPI downloads of mcp in a month (to Oct 1, 2026) · PyPI Stats
- 24k GitHub stars, Python SDK · GitHub
A high CVE count can mean heavy scrutiny as much as weak engineering, so read the incidents and advisories, not only the number. The counts are what this center has recorded, not every flaw ever found; usage figures are the vendors' and registries' own published numbers.
Sources
- npm download counts: npm downloads of @modelcontextprotocol/sdk in a week (Sep 23 to 29, 2026)
- PyPI Stats: PyPI downloads of mcp in a month (to Oct 1, 2026)
- GitHub: GitHub stars, Python SDK
- GitHub Advisory Database: @modelcontextprotocol/sdk
- GitHub Advisory Database: mcp
- GitHub Advisory Database: mcp-remote
- GitHub Advisory Database: @modelcontextprotocol/inspector
- GitHub Advisory Database: @modelcontextprotocol/server-filesystem
- GitHub Advisory Database: mcp-server-kubernetes
- GitHub Advisory Database: figma-developer-mcp
- GitHub Advisory Database: @mcpjam/inspector
- Corelight: The supply-chain risk hiding in your MCP servers
- Help Net Security: Your pinned MCP server may not be as pinned as you think
- Symantec: AI LLMjacking, Skynet
- Help Net Security: hackers hijack exposed AI models to launch cyberattacks at victims' expense
- SOCRadar: INC and The Gentlemen ransomware AI infrastructure
- Pillar Security: Open and shut, remote code execution in OpenCode via MCP OAuth
- OpenCode v1.18.33 release
- CSO Online: hallucinated domain in official OpenAI skill underscores AI supply chain risks
- AIR Security: the Render skill in OpenAI's repository
- Langflow: GitHub advisory GHSA-7w94-79vh-5mr2, authenticated RCE via MCP stdio transport
- Previdian: CVE-2026-105740 alert and honeypot telemetry
- The Hacker Wire: CVE-2026-105697
- BleepingComputer: PoeLLM malware infects exposed AI servers in cryptomining attacks