On October 22, 2026, a contributor to the Model Context Protocol’s specification repository posted a narrowed version of SEP-1766, a digest-pinning proposal first opened in November 2025. The mechanism is modest. The client computes a SHA-256 digest over each tool definition as the server hands it over, then re-checks that digest on every tools/list.1 The next day the proposal’s original author agreed to co-author.2 That is where MCP’s supply-chain story stands as of this writing: almost a year after the first digest-pinning proposal opened, the integrity work consists of two people agreeing to write a document, and the specification revision that shipped in July 2026 has nothing in its changelog about signing, provenance, or attestation.3
I think that is worth stopping on, because npm already ran this experiment. The original sin of package registries was trust by default. Acquiring code and running it collapsed into a single action, and for years the only thing standing between a developer and a malicious post-install script was the hope that nobody had bothered. The industry eventually built a real answer, and the answer has a name: provenance. The Model Context Protocol, the wiring agent tools now use to reach the outside world, has rebuilt the trust-by-default model without that answer. And the part I want to argue is harder than “they forgot.” MCP’s design makes the answer difficult to retrofit. You can sign an artifact. You cannot sign a conversation.
The lesson npm already learned
For most of its life, npm install was an act of faith. A package could declare an install or postinstall script, and that script ran with your user’s privileges as part of pulling the dependency down. The threat model was never subtle. It was the foundation. Lockfiles helped, but a lockfile pins what you installed by content hash; it says nothing about whether that content is what the maintainer actually built, from the source they claim, on infrastructure they control. A pinned hash of poisoned code is still poisoned code.
The fix that finally addressed the right question was build provenance. In April 2023, npm shipped package provenance in public beta, backed by Sigstore. It ties a published package to the source repository, commit, and CI workflow that produced it, using the build system’s OIDC identity to sign the attestation.4 Sigstore itself, which pairs short-lived signing certificates with a public, append-only transparency log,5 had reached general availability the previous October under the OpenSSF.6 PyPI moved in parallel: Trusted Publishing in April 2023 to kill long-lived upload tokens,7 then PEP 740 digital attestations in November 2024 to bind a published file to its source and build.8 None of this is perfect, and I would not call adoption universal. But the primitive exists, and it answers the question lockfiles cannot.
Provenance works because of two preconditions that are easy to overlook. First, there is an immutable, content-addressable artifact: a tarball or a wheel, hashed, frozen, the same bytes for everyone who downloads it. That is the thing you sign. Second, there is a verifier positioned before execution, whether the install client or a human reading the little “published with provenance” badge, that can check the attestation while there is still time to refuse. Hold onto both. MCP has neither.
MCP kept the trust and dropped the seatbelts
Start with what the MCP specification says about its own security posture. The honest sentence is right there in the current revision: “While MCP itself cannot enforce these security principles at the protocol level,” it states, before handing the entire problem to implementers and users via consent dialogs.9 The trust model is not weak by accident. It is, by the protocol’s own admission, not the protocol’s job.
That would be tolerable if the consent dialog were looking at the dangerous part. It isn’t. MCP tools are, in the spec’s word, “model-controlled”: the model reads each tool’s description and annotations and decides on that basis whether and how to call it. Those descriptions go straight into the model’s context. The spec even concedes the danger. It instructs that clients “MUST consider tool annotations to be untrusted unless they come from trusted servers.”10 That is circular. A trustworthy way to establish that a server is trusted is exactly the thing missing. The human approves a server by name in a dialog box. The model consumes a paragraph of natural-language instructions the human never reads. The trust boundary and the attack surface are not the same surface.
The incidents of the last year are symptoms of that gap, not freak events. JFrog found a command-injection flaw in mcp-remote, the proxy that lets local clients reach remote MCP servers, that let a malicious server run arbitrary commands on the client. It became CVE-2025-6514, which JFrog scored 9.6 on CVSS 3.1.11 JFrog’s write-up traces the injection to the authorization_endpoint URL a malicious server hands back during OAuth setup.12 A package called postmark-mcp posed as Postmark’s official MCP server. For fifteen releases it worked as advertised. Then version 1.0.16 added one line that blind-copied every outgoing email to an outside address, and Koi Security caught it.13 Nothing in that package was “vulnerable” in the traditional sense. The code did precisely what its author intended, which is the whole point. Invariant Labs had already named the underlying class in April 2025: tool-poisoning attacks, where malicious instructions live inside a tool description, plus the “rug pull,” where a server serves a benign tool definition until after you have approved it and then changes it.14
Notice where an MCP server’s identity actually lives. Under the spec’s two standard transports, a server is either a subprocess the client launches from a shell command (npx -y some-server, say) or an HTTP endpoint the client posts messages to.15 In the first case, the supply chain collapses straight back into npm or PyPI or a container registry, the same ecosystems we were just talking about. In the second, identity collapses into DNS and TLS, which attest that you reached the right host and say nothing about what code that host runs. MCP added a model-facing instruction layer on top of those ecosystems and inherited all of their trust problems, plus a new one of its own.
You can’t sign a conversation
This is where the npm playbook stops porting cleanly, and it is the part most “MCP needs supply-chain security” takes skip past. Recall provenance’s two preconditions. MCP breaks both, and it breaks them by design rather than by oversight.
There is no immutable artifact to sign. A server advertises its capabilities through the tools/list method. The spec says that set may change over time, and a server that declares the listChanged capability can emit notifications/tools/list_changed to any client listening and hand back a different set of tools on the next call.16 The thing you would want to sign is not a frozen file; it is a live stream of declarations that the server is free to revise mid-session. Whatever a user approved at the start can be replaced afterward. The rug pull is not a clever exploit of MCP. It is a documented protocol feature used against its operator.
And there is no verifier in the right place. Trail of Bits described the timing precisely in their work on “line jumping”: a server’s tool descriptions enter the model’s context the instant the client calls tools/list, which happens before any tool is invoked and before any human approves anything.17 A signature on “the server” would gate the wrong moment. The malicious description has already reached the model and can already steer it. Trail of Bits put the consequence bluntly, calling it the collapse of human-in-the-loop into “human-as-the-rubber-stamp.”18 Even if every tool manifest were signed, the entity actually reading the descriptions is the model, and the model does not check signatures. The one party that could verify provenance has been relocated to an approve-once dialog upstream of the stream it was supposed to watch.
The people closest to the protocol know this. SEP-1766, as opened in November 2025, asked servers to publish a SHA-256 digest for every tool version alongside the tool listing, and it explicitly reserved signatures, provenance, and revocation for “future phases.”19 It never found a sponsor. On June 24, 2026 it was closed administratively, because SEPs had moved to pull requests.20 The October 2026 successor is narrower and, I think, more honest about where verification has to happen: the client digests the tool definition it actually received and re-checks on every tools/list.21 That catches drift. It cannot tell you the first definition was clean, and nothing in it is provenance; it is a tripwire on a stream the protocol lets the server rewrite. As far as I can find, it is the only integrity proposal moving. The MCP Registry that launched in September 2025 lists servers on self-reported metadata and relies on the community to flag malicious or impersonating entries after the fact; it does not attest to the code a listing serves.22 The roadmap, last updated August 22, 2026, does not mention provenance at all. Its security priority is agent identity: DPoP, delegation, and token exchange.23 The gap is not unacknowledged. It is structural.
The agent is the new post-install script
Here is where “at agent speed” stops being a tagline. npm’s post-install script ran the moment you installed. The agent is now the acquire-and-run step fused into one, minus the human pause that used to sit, however briefly, between deciding to pull something and watching it execute.
Miasma showed the fusion in June 2026. The worm, which carries the Shai-Hulud lineage in its own payload, did not wait for npm install. SafeDep found it wired into Claude Code and Gemini CLI SessionStart hooks, an always-applied Cursor rule, and a VS Code task set to run on folder open: code that executes the moment an agent attaches to the repository. As SafeDep put it, “Cloning the repo is safe. Opening it is not.”24 GitHub disabled 73 Microsoft repositories across four of the company’s organizations in response.25 The s1ngularity attack against Nx, in August 2025, had already shown the next move: turning the agent into a weapon rather than just a trigger. Its post-install malware tried to use the developer’s installed AI CLIs, Claude and Gemini among them, as part of its sweep for sensitive data.26 Wiz documented how: it prompted those tools with --dangerously-skip-permissions, --yolo, and --trust-all-tools to pull filesystem contents in its hunt for secrets.27 GitGuardian counted 2,349 distinct secrets leaked across more than a thousand repositories.28 The supply chain didn’t just run code on the victim’s machine. It borrowed the victim’s agent to go looking for more.
Speed is the other half. When LiteLLM was compromised in March 2026, two malicious versions were downloaded 46,996 times before PyPI quarantined them, roughly 46 minutes after they went up. Of the 2,337 packages depending on litellm, about 88% had version specifications loose enough to accept the bad release.29 The blast radius pulled in transitive dependents including Microsoft GraphRAG, Google ADK, DSPy, and OpenHands.30 Forty-six minutes is not enough time for a human to notice, let alone intervene. In an agentic install graph, the window between a package going bad and that package executing is shrinking toward zero, and the agent is what’s closing it.
The controls agents actually ship with don’t help, because they aren’t provenance controls. Cursor’s auto-run allowlist is the seatbelt that is supposed to keep an agent from running arbitrary commands. Before version 2.3, certain shell built-ins ran without appearing in the allowlist and without approval, which let a prompt-injected agent set or alter environment variables that the allowlisted, trusted commands then inherited.31 That was CVE-2026-22708, scored 9.8 on NVD’s CVSS 3.1 scale.32 The allowlist checks the name of the command. It never asks what environment that command will run in, or where its binary came from. That is the whole problem in miniature: our agent guardrails reason about identity and permission, and almost never about origin.
What I think survives
This is opinion, stated plainly. Provenance is the missing primitive, and I don’t think that’s a controversial claim. The uncomfortable part is the next one: MCP cannot get there by copying npm, because npm’s fix assumes an immutable artifact and a verifier that reads the attestation before the code runs, and MCP, by design, offers neither. Waiting for a protocol-level Sigstore is waiting for something the protocol’s shape resists.
So the workable defenses, in my view, are the ones that stop pretending the artifact model applies. The most promising thing I’ve seen is client-side and boring: trust-on-first-use pinning of tool descriptions and server instructions, with drift detection that blocks a tool whose definition changed after you approved it. Trail of Bits shipped exactly that in mcp-context-protector,33 and the October proposal in the spec repository would make the same check a protocol behavior rather than a wrapper.34 That is not signing the stream; it is noticing when the stream changes, which is the most you can do to something that won’t hold still. Pair it with tool surfaces scoped tightly enough that a poisoned tool can’t reach anything that matters, and keep a human at the execute boundary rather than the approve-once boundary, where the agent is about to do something it can’t take back.
None of that is as clean as a green provenance badge, and none of it will arrive bundled in a quickstart. npm took the better part of a decade to internalize that acquiring code and trusting it are different acts. The agents are repeating the lesson at a speed that doesn’t grant a decade.
Notes
Footnotes
-
Comment by gautamgb on SEP-1766, “Digest-Pinned Tool Versioning and Interceptor-Based Validation in MCP,” modelcontextprotocol/modelcontextprotocol issue #1766, October 22, 2026. Proposes that the client compute a SHA-256 digest over each tool definition as received and re-check it on every
tools/list. https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1766 ↩ -
Comment by ev3rl0ng, the SEP-1766 author, on the same issue, October 23, 2026, agreeing to co-author the narrowed proposal. https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1766 ↩
-
“Key Changes,” Model Context Protocol specification, revision 2026-07-28. The changelog lists no change concerning signing, provenance, or attestation. https://modelcontextprotocol.io/specification/2026-07-28/changelog ↩
-
Brian DeHamer and Philip Harrison, “Introducing npm package provenance,” The GitHub Blog, April 19, 2023 (public beta). https://github.blog/security/supply-chain-security/introducing-npm-package-provenance/ ↩
-
“Overview,” Sigstore documentation, undated (accessed October 22, 2026). Describes the short-lived certificate issued by Fulcio and the append-only Rekor transparency log. https://docs.sigstore.dev/about/overview/ ↩
-
“Sigstore Announces General Availability at SigstoreCon,” OpenSSF, October 25, 2022. https://openssf.org/press-release/2022/10/25/sigstore-announces-general-availability-at-sigstorecon/ ↩
-
Dustin Ingram, “Introducing Trusted Publishers,” The Python Package Index Blog, April 20, 2023. https://blog.pypi.org/posts/2023-04-20-introducing-trusted-publishers/ ↩
-
Dustin Ingram, “PyPI now supports digital attestations,” The Python Package Index Blog, November 14, 2024, implementing PEP 740. https://blog.pypi.org/posts/2024-11-14-pypi-now-supports-digital-attestations/ ↩
-
“Security and Trust & Safety,” Model Context Protocol specification, revision 2026-07-28. Quoted: “While MCP itself cannot enforce these security principles at the protocol level.” https://modelcontextprotocol.io/specification/2026-07-28 ↩
-
“Tools,” Model Context Protocol specification, revision 2026-07-28. Tools are described as “model-controlled”; the spec instructs that clients “MUST consider tool annotations to be untrusted unless they come from trusted servers.” https://modelcontextprotocol.io/specification/2026-07-28/server/tools ↩
-
CVE-2025-6514, OS command injection in
mcp-remotewhen connecting to untrusted servers. NVD record published July 9, 2025; the CVSS 3.1 base score of 9.6 listed there is supplied by JFrog as the reporting CNA. https://nvd.nist.gov/vuln/detail/CVE-2025-6514 ↩ -
Or Peles, “Critical RCE Vulnerability in mcp-remote: CVE-2025-6514 Threatens LLM Clients,” JFrog Security Research, July 9, 2025. Vendor research blog. https://jfrog.com/blog/2025-6514-critical-mcp-remote-rce-vulnerability/ ↩
-
Idan Dardikman, “First Malicious MCP in the Wild: The Postmark Backdoor That’s Stealing Your Emails,” Koi Security, September 25, 2025. Vendor research blog. The original URL now redirects to an unrelated product page, so the citation points at the Internet Archive snapshot of October 5, 2025, which records versions 1.0.0 through 1.0.15 as clean and the BCC added in 1.0.16. https://web.archive.org/web/20251005135359/https://www.koi.ai/blog/postmark-mcp-npm-malicious-backdoor-email-theft ↩
-
Luca Beurer-Kellner and Marc Fischer, “MCP Security Notification: Tool Poisoning Attacks,” Invariant Labs, April 1, 2025. https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks ↩
-
“Transports,” Model Context Protocol specification, revision 2026-07-28, defining the stdio binding (newline-delimited messages over the standard streams of a client-launched subprocess) and Streamable HTTP (each message an HTTP POST to a single MCP endpoint). https://modelcontextprotocol.io/specification/2026-07-28/basic/transports ↩
-
“Tools,” Model Context Protocol specification, revision 2026-07-28. The tool set “MAY change over time”; servers declaring the
listChangedcapability sendnotifications/tools/list_changedto subscribed clients, which then re-fetch viatools/list. https://modelcontextprotocol.io/specification/2026-07-28/server/tools ↩ -
“Jumping the line: How MCP servers can attack you before you ever use them,” Trail of Bits, April 21, 2025. https://blog.trailofbits.com/2025/04/21/jumping-the-line-how-mcp-servers-can-attack-you-before-you-ever-use-them/ ↩
-
Same Trail of Bits post, April 21, 2025. Quoted: the model “effectively transforms the ‘human-in-the-loop’ security model into ‘human-as-the-rubber-stamp.’” https://blog.trailofbits.com/2025/04/21/jumping-the-line-how-mcp-servers-can-attack-you-before-you-ever-use-them/ ↩
-
ev3rl0ng, SEP-1766, “Digest-Pinned Tool Versioning and Interceptor-Based Validation in MCP,” issue #1766, opened November 5, 2025. Quoted from the proposal: “All additional metadata (signatures, provenance, SBOM, revocation) is OPTIONAL and reserved for future phases.” https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1766 ↩
-
Closure of issue #1766, June 24, 2026, noting that SEPs had moved to pull requests. https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1766 ↩
-
Comment by gautamgb on issue #1766, October 22, 2026, describing the client-side digest re-checked on every
tools/list. https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1766 ↩ -
David Soria Parra, Adam Jones, Tadas Antanavicius, Toby Padilla, and Theodora Chu, “Introducing the MCP Registry,” Model Context Protocol Blog, September 8, 2025. Listings are described as self-reported; moderation runs through community-filed issues flagging spam, malicious code, or impersonation. https://blog.modelcontextprotocol.io/posts/2025-09-08-mcp-registry-preview/ ↩
-
“Roadmap,” modelcontextprotocol.io, last updated August 22, 2026. Priority area 3, “Agent Identity and Enterprise-Ready Security,” covers DPoP, agent identity and delegation, and RFC 8693 token exchange. The page does not mention provenance. https://modelcontextprotocol.io/development/roadmap ↩
-
SafeDep, “Miasma Worm Targets AI Coding Agents via GitHub Repos,” June 5, 2026. Vendor research blog. https://safedep.io/miasma-worm-ai-coding-agent-config-injection/ ↩
-
Ravie Lakshmanan, “Miasma Worm Hits 73 Microsoft GitHub Repositories in Major Supply Chain Attack,” The Hacker News, June 6, 2026. Secondary reporting; the 73-repository count is attributed there to OpenSourceMalware. https://thehackernews.com/2026/06/miasma-worm-hits-73-microsoft-github.html ↩
-
Juri Strumpflohner, “S1ngularity: What Happened, How We Responded, What We Learned,” Nx Blog, September 5, 2025. https://nx.dev/blog/s1ngularity-postmortem ↩
-
Merav Bar and Rami McCarthy, “s1ngularity: supply chain attack leaks secrets on GitHub: everything you need to know,” Wiz, August 27, 2025. Vendor research blog. https://www.wiz.io/blog/s1ngularity-supply-chain-attack ↩
-
Guillaume Valadon and Anna Nabiullina, “The Nx ‘s1ngularity’ Attack: Inside the Credential Leak,” GitGuardian, August 27, 2025. Vendor research blog. 2,349 distinct secrets; 1,079 repositories contained at least one. https://blog.gitguardian.com/the-nx-s1ngularity-attack-inside-the-credential-leak/ ↩
-
Daniel Hnyk, “LiteLLM Hack: Were You One of the 47,000?,” FutureSearch, March 25, 2026. 46,996 downloads of versions 1.82.7 and 1.82.8; quarantined within 46 minutes; 2,337 dependent packages, of which 2,054 (88%) had version specs that accepted the compromised releases. https://futuresearch.ai/blog/litellm-hack-were-you-one-of-the-47000/ ↩
-
Sai Likhith, “litellm: Credential Stealer Hidden in PyPI Wheel,” StepSecurity, March 24, 2026. Vendor research blog. https://www.stepsecurity.io/blog/litellm-credential-stealer-hidden-in-pypi-wheel ↩
-
GHSA-82wg-qcm4-fp2w, “Terminal Tool Allowlist Bypass via Environment Variables,” Cursor security advisory, January 14, 2026. Affects versions through 2.2; fixed in 2.3. https://github.com/cursor/cursor/security/advisories/GHSA-82wg-qcm4-fp2w ↩
-
CVE-2026-22708, NVD record published January 14, 2026. NVD CVSS 3.1 base score 9.8; the CNA separately assigned a CVSS 4.0 base score of 7.2. https://nvd.nist.gov/vuln/detail/CVE-2026-22708 ↩
-
Cliff Smith, “We built the security layer MCP always needed,” Trail of Bits, July 28, 2025. Introduces
mcp-context-protector; a changed tool description is blocked from the downstream app, and a change to server instructions blocks the whole server, until re-approved. https://blog.trailofbits.com/2025/07/28/we-built-the-security-layer-mcp-always-needed/ ↩ -
Comment by gautamgb on issue #1766, October 22, 2026. https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1766 ↩


