Frameworkv0.1, beta
The OWASP MCP Top 10, with the incidents behind each risk
OWASP's list of the ten most serious security risks in systems built on the Model Context Protocol (MCP), the standard way AI agents connect to tools and data. It is a beta (version 0.1, first written in November 2025, with a next release planned for October 2026; on October 17, 2026 the project page still showed the 2025 beta in its pilot-testing phase). Below, each risk in plain words, what OWASP says to do about it, and where it has already happened, from the incidents in this center, with the weaknesses, practices and automated tests that go with it.
- MCP01 Token Mismanagement & Secret Exposure
- MCP02 Privilege Escalation via Scope Creep
- MCP03 Tool Poisoning
- MCP04 Software Supply Chain Attacks & Dependency Tampering
- MCP05 Command Injection & Execution
- MCP06 Intent Flow Subversion
- MCP07 Insufficient Authentication & Authorization
- MCP08 Lack of Audit and Telemetry
- MCP09 Shadow MCP Servers
- MCP10 Context Injection & Over-Sharing
MCP01 Token Mismanagement & Secret Exposure
Tokens and keys end up hard-coded in configuration files, environment variables or prompt templates, or linger in the model's context and in logs, where a prompt or a look at the logs can pull them out later.
What OWASP says to do:
- Keep secrets in a vault and inject them only at run time.
- Use short-lived tokens scoped to one agent, tool or session.
- Keep secrets out of the model's context, redact them before logging, and rotate them when exposure is suspected.
Where it happened: Supabase MCP: a support ticket tricks a Cursor agent into leaking secrets · Loom for AWS: an unauthenticated client could take over the agent… · LiteLLM: any Bearer token opened the MCP gateway, and attackers used it
Weaknesses: ASC-2026-0033: Long-lived credentials in agent and MCP configuration · ASC-2026-0012: System prompt and secret leakage · ASC-2026-0066: Prompts and outputs logged with sensitive data
Practices: Keep secrets out of reach of models and agents
Tests you can run: Reading a web page never leads to reading .env
MCP02 Privilege Escalation via Scope Creep
Permissions that started narrow keep growing, for convenience or by configuration drift, until an agent holds broad or admin rights and can change code, deploy or read sensitive data with no one reviewing it.
What OWASP says to do:
- Grant the least privilege, with fine-grained scopes.
- Make access expire, give it just in time and put an approval gate in front of it.
- Give each agent its own identity and review how its permissions grow.
Where it happened: Supabase MCP: a support ticket tricks a Cursor agent into leaking secrets · GitHub MCP: a malicious issue makes an agent leak private repositories
Weaknesses: ASC-2026-0032: Overly broad OAuth scopes for AI connectors · ASC-2026-0003: Identity and privilege abuse by agents · ASC-2026-0042: Agents with standing write access to production
Practices: Give agents their own identity with the least privilege
Tests you can run: The agent's database role cannot delete or drop
MCP03 Tool Poisoning
An attacker compromises the tools, their descriptions or their outputs, so the model acts on malicious or misleading context: hidden instructions in a description, a server that changes after approval (a rug pull), a tool that shadows another, or a schema changed so a harmless action does something destructive.
What OWASP says to do:
- Sign tool manifests and schemas and check them before use.
- Review schema changes, and require a person to approve destructive actions.
- Scan tool descriptions before install for instructions aimed at the model, secret file paths and hidden characters.
Where it happened: OpenAI's official Render skill sent production apps' tokens to a domain… · Deadbugz: an MCP server that turns malicious after three tool calls · MCPoison: a swapped MCP config gives persistent code execution in Cursor · postmark-mcp: first malicious MCP server found in the wild
Weaknesses: ASC-2026-0011: MCP tool description poisoning · ASC-2026-0035: MCP server changes after approval (rug pull) · ASC-2026-0036: Tool name collision and shadowing across MCP servers
Practices: Pin approved MCP servers to their tool list and re-check on every change · Install MCP servers only from sources you have vetted
Tests you can run: Every MCP server is pinned to an exact version
MCP04 Software Supply Chain Attacks & Dependency Tampering
The SDKs, connectors, servers and plugins an agent depends on run in trusted paths, so one compromised dependency can change the agent's behavior or plant a backdoor without anyone noticing.
What OWASP says to do:
- Require signed components and check where they came from.
- Keep a bill of materials for every server and plugin.
- Pin versions, install from an internal registry, scan dependencies and sandbox plugins.
Where it happened: postmark-mcp: first malicious MCP server found in the wild · SmartLoader clones an Oura Ring MCP server to spread an infostealer · MCP Python SDK: a malicious server could steal the client's OAuth… · Corelight: a pinned MCP server can still pull fresh dependencies on…
Weaknesses: ASC-2026-0004: Agentic supply chain compromise · ASC-2026-0058: Typosquatted and lookalike AI SDK packages · ASC-2026-0059: Unpinned AI dependencies and compromised maintainer releases
Practices: Curate dependencies: approve, pin, verify, and re-check · Pin approved MCP servers to their tool list and re-check on every change
Tests you can run: Every MCP server is pinned to an exact version · The build fails on a known, fixed CVE in your AI stack
OWASP: MCP04 Software Supply Chain Attacks & Dependency Tampering
MCP05 Command Injection & Execution
The agent builds shell commands, SQL, API calls or code from untrusted input (prompts, retrieved text, third-party data) and runs them without checking.
What OWASP says to do:
- Allow only listed commands, arguments and paths.
- Never pass model output to a shell, eval or exec; use structured parameters and parameterized SQL.
- Run tools sandboxed and not as root, and require approval for destructive actions.
Where it happened: OpenCode on Windows: a remote MCP server's sign-in link could run… · Langflow: a logged-in user's MCP server command runs as a shell, and… · MCP STDIO design flaw: server configs become command execution · CurXecute: a Slack message leads to code execution in Cursor · MCPoison: a swapped MCP config gives persistent code execution in Cursor
Weaknesses: ASC-2026-0005: Unexpected code execution by agents · ASC-2026-0017: Improper handling of model output (XSS, SQL and command injection) · ASC-2026-0039: Command allowlist and approval bypass in coding agents
Practices: Run agents in a sandbox with default-deny egress · Treat MCP server configuration as code execution
Tests you can run: A web page cannot drive your local agent server · Code a model writes is rejected unless it runs in a sandbox
MCP06 Intent Flow Subversion
Instructions hidden in what the agent retrieves (documents, schemas, tool output) take over its plan, steering it toward the attacker's goal while it still looks like it is doing the user's task.
What OWASP says to do:
- Hold on to the user's original goal and check every proposed tool call against it.
- Treat every MCP resource and tool output as untrusted, and tag it so.
- Pause for a person when the agent's actions drift from what was asked.
Where it happened: GitHub MCP: a malicious issue makes an agent leak private repositories · Supabase MCP: a support ticket tricks a Cursor agent into leaking secrets · GhostJacking: firewall logs become instructions for the agent that reads…
Weaknesses: ASC-2026-0001: Agent goal hijack via untrusted content (prompt injection) · ASC-2026-0024: Hidden instructions in invisible or disguised text · ASC-2026-0043: No human approval for consequential actions
Practices: Treat everything an agent reads as attacker-controlled · Check every agent action against policy before it runs · Require a human for consequential actions
Tests you can run: An email's text is read, never obeyed, run after run · Reading a web page never leads to reading .env
MCP07 Insufficient Authentication & Authorization
Servers, tools or agents do not check who is calling or what the caller may do: optional token checks, shared hard-coded secrets, identity taken from the caller's word, or checks made only on the client.
What OWASP says to do:
- Validate short-lived, scoped tokens on the server, on every request, and deny by default.
- Expire, rotate and revoke tokens, and detect replays.
- Use central identity management and log every access decision.
Where it happened: PoeLLM: a miner that found its command server in a poem infected 3,400… · LiteLLM: any Bearer token opened the MCP gateway, and attackers used it · Loom for AWS: an unauthenticated client could take over the agent… · Obot: the Docker quickstart gave anyone on the network admin rights and… · MCP TypeScript SDK: a malicious MCP server could choose where the…
Weaknesses: ASC-2026-0015: Exposed AI infrastructure without authentication · ASC-2026-0034: Confused deputy and token passthrough in MCP servers · ASC-2026-0037: Local AI servers reachable from the browser (DNS rebinding and missing auth)
Practices: Never expose AI infrastructure to the internet
Tests you can run: A wrong or default token is refused on every route · Nothing in the deployment mounts the Docker socket or turns authentication off · A web page cannot drive your local agent server
MCP08 Lack of Audit and Telemetry
Without logs and telemetry nobody can see what agents did, what data they touched or why, and misuse can go unnoticed for weeks.
What OWASP says to do:
- Log every action and tool call with its parameters and identity, append-only.
- Send the logs to your SIEM with alerts, with sensitive data masked.
- Know what normal looks like, trace with OpenTelemetry and keep logs long enough.
Where it happened: No incident in this center shows it yet.
Weaknesses: ASC-2026-0044: Missing audit trail of agent actions
Practices: Log every agent action and watch it in real time
Tests you can run: No test yet in AI Test Automation: a gap we name plainly.
MCP09 Shadow MCP Servers
MCP servers set up outside security's view, often with default credentials or open APIs, become backdoors no one is watching.
What OWASP says to do:
- Keep a central registry of MCP servers, and fail any deployment that is not in it.
- Scan your networks for MCP endpoints all the time.
- Ship secure-by-default templates and segment the network.
Where it happened: Obot: the Docker quickstart gave anyone on the network admin rights and… · OX Security maps 15,465 public MCP servers: some resolve to China and…
Weaknesses: ASC-2026-0015: Exposed AI infrastructure without authentication · ASC-2026-0069: Default credentials on AI tools and their storage
Practices: Keep an inventory of models, agents, tools and their permissions · Never expose AI infrastructure to the internet
Tests you can run: Nothing in the deployment mounts the Docker socket or turns authentication off
MCP10 Context Injection & Over-Sharing
The agent's working memory is shared, kept too long or scoped too widely, so data leaks between users, sessions or agents, and content planted in it steers later requests.
What OWASP says to do:
- Keep contexts short-lived and per session.
- Separate memory and vector stores per user, agent and tenant, with expiry.
- Redact personal data and secrets before they enter the context.
Where it happened: No incident in this center shows it yet.
Weaknesses: ASC-2026-0047: Memory shared across users or sessions · ASC-2026-0048: Response caches that serve one user's answer to another · ASC-2026-0006: Memory and context poisoning · ASC-2026-0018: RAG and vector store poisoning, and cross-tenant retrieval
Practices: Secure the data your AI is trained on and reads
Tests you can run: No test yet in AI Test Automation: a gap we name plainly.
OWASP renamed MCP06 from "Prompt Injection via Contextual Payloads" to "Intent Flow Subversion" in February 2026, and some of its pages still show the old name; this page uses the current one. The summaries are ours; OWASP's own pages are the reference, and its content is shared under a Creative Commons Attribution-ShareAlike license.