Anthropic shipped the Model Context Protocol in November 2024 with an easy pitch: a universal connector, the USB-C of AI agents. Two years later the US National Security Agency is publishing design-consideration guidance warning that MCP clients routinely fail to resolve tool names safely, and the Cloud Security Alliance calls MCP "one of the most rapidly weaponized attack surfaces in agentic AI." The ecosystem replayed, in half the time, the supply chain lessons the JavaScript world spent a decade learning. The difference is that these dependencies do not just run code. They talk to your model.
Two years of incidents, compressed
The record reads like a greatest hits of package-registry failures, with an LLM twist.
- April 2025. Invariant Labs documented tool poisoning: malicious instructions buried in an MCP tool description, visible to the model but not rendered in the client UI. In a live demo, a trivia-game server hijacked an agent session and exfiltrated message history through an adjacent, trusted WhatsApp MCP server. Simon Willison framed the same week what the protocol had structural trouble with: untrusted text landing in a trusted context.
- May 2025. The same team showed a crafted GitHub issue could make an agent with GitHub MCP access read private repositories and leak them into public pull requests. A legitimate tool call, steered by injected text.
- July 2025. The Supabase MCP server became the shorthand example of what Willison called the lethal trifecta: private data, untrusted content, and an outbound channel in one agent. A scan that month found at least 1,862 MCP servers on the public internet answering requests with no authentication at all; the authorization spec marks OAuth 2.1 support as optional.
- August 2025. Cymulate disclosed sandbox escape and symlink containment flaws (CVE-2025-53109, CVE-2025-53110) in Anthropic's own Filesystem MCP server.
- September 2025. The first confirmed in-the-wild malicious MCP server: a package impersonating a Postmark email integration. Per the CSA, it shipped fifteen clean versions to npm to build trust, then added a single line in version 1.0.16 that blind-carbon-copied every email through the agent.
- October 2025. GitGuardian found a path traversal in the smithery.yaml build config of the Smithery hosting registry. A crafted build could leak the builder's Docker credentials, including a Fly.io API token controlling more than 3,000 hosted MCP servers, and tap inbound client traffic carrying API keys. The registry itself became the compromise.
- January to March 2026. A zero-day command injection (CVE-2026-0755, CVSS 9.8) in the unofficial gemini-mcp-tool; SmartLoader operators cloning a legitimate Oura ring MCP project, complete with fake contributors and fork farms, to deliver StealC infostealers through public registries; and CVE-2026-33032 in nginx-ui, where an unauthenticated MCP endpoint allowed full service takeover on more than 2,600 exposed instances.
- April 2026. OX Security published "The Mother of All AI Supply Chains": command execution built into the STDIO transport of every official MCP SDK (Python, TypeScript, Java, Rust). Configuration parameters pass to the host shell unsanitized, so an attacker who can influence an MCP config file gets arbitrary code execution. The firm estimated 200,000 vulnerable instances across a supply chain with more than 150 million package downloads, and the disclosure produced more than 30 advisories and at least seven high- or critical-severity CVEs across MCP Inspector, Cursor, LibreChat, LiteLLM and others, with Windsurf still unpatched as of May 2026 (CVE-2026-30615).
The vendor response to the STDIO finding is the part security teams remembered: Anthropic confirmed the behavior is intentional, declined architectural changes, and updated SECURITY.md to note that STDIO adapters should be used with care. Remediation responsibility moved downstream.

Why this is worse than npm
MCP inherited npm's failure modes: impersonation, typosquatting, maintainer compromise, automatic-update rug pulls. A systematic study of malicious MCP servers catalogued twelve attack classes against hosts and users, including servers configured with update flags that silently refetch modified packages after a user has already approved them.
But three properties make the blast radius different from ordinary dependencies.
First, MCP servers are dependencies you grant credentials to. The Postmark server sat between an agent and a mailbox; the GitHub incident needed no exploit, only an over-scoped personal access token and a hostile issue. The CSA note on tool poisoning frames this as an attack on "the capability supply chain that agents depend on," now codified as MCP03:2025 in the OWASP MCP Top 10.
Second, an MCP dependency has two attack surfaces: its code and its text. A package can be audited and safe at install time while its tool description, a string the model reads and the user rarely sees, carries instructions to call other tools. Microsoft's guidance describes exactly this asymmetry, and cross-server shadowing means one poisoned server can misdirect the agent's use of a legitimate neighbor. Unit 42 added the sampling vector: malicious servers asking the client's LLM to generate text back into the conversation, persisting injected instructions across turns.
Third, the trust unit is a plain JSON config. Pasting a server block from a blog or README into a client grants it whatever the user's agent can touch. Kaspersky GERT demonstrated a server that looked routine and harvested developer credentials on every tool call.
What the mitigations actually are
Nothing invented for MCP is exotic. The ecosystem arrived, expensively, at standard supply chain hygiene plus disclosure.
- Pin and hash. Registries and clients should pin server versions and verify hashes, re-prompting users when tool definitions change; the ETDI proposal and Red Hat both center on preventing silent definition drift, the rug pull.
- Treat every server as an untrusted third party. The CSA's headline recommendation for enterprises: zero-trust controls at the tool layer, authentication on by default, and governance as an ongoing program rather than a patch cycle.
- Scope credentials to the tool, not the user. The GitHub and Supabase incidents were least-privilege failures, not parser bugs.
- Render tool descriptions to humans. Several client vendors, Cursor among them after Check Point's MCPoison disclosure, moved to surface full descriptions and diff them on change.
- Sanitize at the transport. OX's practical advice: allowed-command lists or explicit dangerous-mode flags instead of passing config strings to a shell.
The uncomfortable summary is that the protocol ships defaults optimized for demo velocity, and safety lands in downstream implementations that most users never inspect. MCP is now the integration layer under Claude Desktop, Cursor, Windsurf, LibreChat and a large part of the agent tooling stack. It has 2015-era npm security maturity with 2026-era privilege grants. The incidents above are not an anomaly in that gap; they are what the gap produces.

Sources
- Anthropic, "Introducing the Model Context Protocol" (November 2024) - https://www.anthropic.com/news/model-context-protocol
- NSA, "Model Context Protocol (MCP): Security Design Considerations" (May 2026) - https://www.nsa.gov/Portals/75/documents/Cybersecurity/CSI_MCP_SECURITY.pdf
- Cloud Security Alliance, "MCP Security Crisis: Systemic Design Flaws in AI Agent Infrastructure" (May 2026) - https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-security-crisis-20260504-csa-styled
- Cloud Security Alliance, "MCP Tool Poisoning: Adversarial Hijacking of AI Agent Workflows" - https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-tool-poisoning-ai-agent-exfiltration-2
- OX Security, "The Mother of All AI Supply Chains: Technical Deep Dive" (April 2026) - https://www.ox.security/blog/the-mother-of-all-ai-supply-chains-technical-deep-dive/
- Infosecurity Magazine, "Systemic Flaw in MCP Protocol Could Expose 150 Million Downloads" - https://www.infosecurity-magazine.com/news/systemic-flaw-mcp-expose-150/
- Invariant Labs, "MCP Security Notification: Tool Poisoning Attacks" (April 2025) - https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks
- Invariant Labs, "Accessing Private Repositories via MCP Prompt Injection" (May 2025) - https://invariantlabs.ai/blog/accessing-private-repositories-via-mcp-prompt-injection-attack
- Simon Willison, "Supabase MCP: The Lethal Trifecta" (July 2025) - https://simonwillison.net/2025/Jul/6/supabase-mcp-lethal-trifecta/
- Dark Reading, "Nearly 2,000 MCP Servers Possess No Security Whatsoever" (July 2025) - https://www.darkreading.com/vulnerabilities-threats/2000-mcp-servers-security
- MCP Authorization Specification - https://modelcontextprotocol.io/specification/2025-03-26/basic/authorization
- Cymulate, "EscapeRoute: Anthropic Filesystem MCP Server" (August 2025) - https://cymulate.com/blog/cve-2025-53109-53110-escaperoute-anthropic/
- GitGuardian, "Breaking MCP Server Hosting" (October 2025) - https://blog.gitguardian.com/breaking-mcp-server-hosting/
- AuthZed, "A Timeline of Model Context Protocol (MCP) Security Breaches" (updated May 2026) - https://authzed.com/blog/timeline-mcp-breaches
- Security Affairs, "SmartLoader Hackers Clone Oura MCP Project to Spread StealC Malware" (February 2026) - https://securityaffairs.com/188135/ai/smartloader-hackers-clone-oura-mcp-project-to-spread-stealc-malware.html
- Microsoft Developer Blog, "Protecting Against Indirect Prompt Injection Attacks in MCP" - https://developer.microsoft.com/blog/protecting-against-indirect-injection-attacks-mcp
- Unit 42, "New Prompt Injection Attack Vectors Through MCP Sampling" - https://unit42.paloaltonetworks.com/model-context-protocol-attack-vectors
- "When MCP Servers Attack: Taxonomy, Feasibility, and Mitigation" (arXiv:2509.24272) - https://arxiv.org/html/2509.24272v1
- Kaspersky Securelist, "Shiny Tools, Shallow Checks: How AI Hype Opens the Door to Malicious MCP Servers" - https://securelist.com/model-context-protocol-for-ai-integration-abused-in-supply-chain-attacks/117473
- OWASP, "MCP Tool Poisoning" - https://community.owasp.org/attacks/MCP_Tool_Poisoning
- Check Point Research, "Cursor IDE's MCP Vulnerability" (2025) - https://research.checkpoint.com/2025/cursor-vulnerability-mcpoison/
- "ETDI: Mitigating Tool Squatting and Rug Pull Attacks in MCP" (arXiv:2506.01333) - https://arxiv.org/html/2506.01333v1
- Red Hat, "Model Context Protocol: Understanding Security Risks and Controls" - https://www.redhat.com/en/blog/model-context-protocol-mcp-understanding-security-risks-and-controls
