CanopiiCanopiiAll serversEnterprise →

October 2026 edition

The MCP Security Index

We independently scanned 43,662 Model Context Protocol servers. 28% scored D or F, 0% ship committed secrets, and 53% declare no authentication at all.

Frozen · captured · scoring rubric v2.11.0 · methodology

Scoring rubric changed between September 2026 (v2.9.0) and this edition (v2.11.0). The deltas below therefore mix real ecosystem movement with a change in how we measure — read them with that in mind.

Graded D or F
28%+10%

12,225 of 43,662 scanned servers carry a failing or near-failing security grade.

Committed secrets
0%no change

Live credentials found in published source — an attacker can use these as-is.

No declared auth
53%+5%

No OAuth 2.1 or protected-resource metadata gating who may invoke the server's tools.

Graded A
32%+4%

13,769 servers verifiably pass the controls that apply to them.

Grade distribution

The median MCP server scores 85/100. Half the corpus falls between 40 and 90.

43,662 scored servers

D/F trend
A13,769 · 32%B14,493 · 33%
C3,061 · 7%
D9,153 · 21%
F3,186 · 7%
12,226 unverifiable
Score distribution summary
Mean score71.9
Median score85
25th percentile40
75th percentile90
Lowest score12
Highest score100

Popularity does not mean safety

Adoption and security posture move independently. The most-installed servers are not the safest ones.

By GitHub stars

CohortserversAverage scoreGraded D/F
1,000+ stars353
78
15%
100–999 stars1,334
79
13%
10–99 stars3,115
80
10%
Under 10 stars27,483
80
10%

By monthly downloads

CohortserversAverage scoreGraded D/F
100k+/month42
83
10%
10k–100k/month146
79
12%
1k–10k/month2,092
85
7%
Under 1k/month11,065
80
18%

Authentication and live exposure

We reached 3,096 live MCP endpoints. Of those, 3% declared that authentication was required and then served their tool surface to an anonymous caller.

Live endpoints probed
3,096

Servers we connected to and spoke MCP with directly.

Auth declared but not enforced
3%

Of probed endpoints — documented auth that anonymous callers get straight past.

Tool-poisoning markers
6%

Hidden instructions found in tool descriptions, which hijack the calling agent.

Attack surface and ecosystem

More exposed tools means more ways to be wrong. Ecosystem mix shows where the risk concentrates.

By number of exposed tools

CohortserversAverage scoreGraded D/F
No tools9,351
82
4%
1–5 tools12,546
71
37%
6–20 tools12,615
72
33%
20+ tools9,150
63
35%

By package ecosystem

CohortserversAverage scoreGraded D/F
repo / remote-only24,278
65
40%
npm14,112
81
15%
pypi4,347
81
8%
mcpb696
84
4%
nuget139
88
5%
cargo90
87
3%

What the ecosystem fails most

Every control in the catalog, ranked by failure rate within each domain. Failure rates are a share of servers where the control actually applied and could be evaluated.

Code safety

ControlFailure rateFailedWarnedEvaluated
No command-injection sinksGuardcode.no_command_injection
2%
65432234,695
No dynamic code executionGuardcode.no_code_eval
1%
35014334,695
No unsafe deserializationGuardcode.no_unsafe_deserialize
0%
962534,695
No path traversalGuardcode.no_path_traversal
0%
09,18434,695
No SSRF sinksGuardcode.no_ssrf
0%
010,94734,695

Secrets & credentials

ControlFailure rateFailedWarnedEvaluated
No committed secretsGuardsecrets.no_committed_secrets
0%
0034,695
Credentials sourced from environmentsecrets.from_env
0%
012,76834,695

Dependencies & supply chain

ControlFailure rateFailedWarnedEvaluated
No install/post-install scriptsGuardsupply.no_install_scripts
3%
407014,109
Package name not typosquattingGuardsupply.not_typosquatting
0%
13018,434
No known-vulnerable dependenciessupply.no_known_vulns
0%
05,11023,360
Dependencies pinned (lockfile)supply.deps_pinned
0%
026,52434,695
Published with provenancesupply.provenance
0%
09,24614,109
Established maintainersupply.maintainer_established
0%
015,61618,434

Tool integrity

ControlFailure rateFailedWarnedEvaluated
No over-broad / destructive toolsGuardtool.no_destructive_scope
8%
2,838034,311
Tool descriptions free of injection markersGuardtool.no_injection_markers
7%
2,488034,311
No risky post-publish tool changes (rug-pull)Guardtool.no_rug_pull
0%
524,06213,530
No in-session tool-definition changesGuardtool.no_runtime_mutation
0%
13624,193
Strict tool input schemastool.schemas_strict
0%
022,53926,361

Auth & transport

ControlFailure rateFailedWarnedEvaluated
Origin validated (DNS-rebinding protection)transport.origin_validated
3%
1222,6903,534
No token passthroughGuardauth.no_token_passthrough
1%
1106213,135
Remote endpoints use TLSGuardtransport.uses_tls
0%
0025,613
Authentication declaredauth.declared
0%
022,98725,613
Execution sandboxingdeploy.sandboxed
0%
028,78034,695
Live endpoint reachabledynamic.reachable
0%
000
Authentication enforceddynamic.auth_enforced
0%
000
No bind-all / exposed debugtransport.no_bind_all
0%
04,33634,695

Maintenance & governance

ControlFailure rateFailedWarnedEvaluated
Repository not archivedGuardmaint.not_archived
1%
364032,285
Actively maintainedmaint.actively_maintained
0%
01,68632,285
Declares a licensegov.declares_license
0%
08,32832,285
Has a security policygov.security_policy
0%
028,27734,695
Signed releasesgov.signed_releases
0%
000
Adoption & popularityreputation.adoption
0%
019,97134,609

How much we could verify

Absence of evidence is not safety. A partially-scannable server cannot present a high score, so confidence caps it.

CohortserversAverage scoreGraded D/F
High (80%+ verified)29,533
81
11%
60–80% verified5,162
78
3%
Low (under 40% verified)8,967
40
100%

A further 12,226 listed servers declare no public repository or their source cannot be retrieved. They are listed without a score rather than given an invented one.

How to read these numbers

How often does the MCP Security Index update?
The live page recomputes from the database continuously; a permanent edition is frozen at the end of every calendar month at /mcp-security-index/YYYY-MM. Servers themselves are re-ingested and re-scanned every six hours, so a score change shows up on the live page the same day.
What does a D or F grade actually mean?
Grades come from a 0–100 score: A is 90 and up, B 75–89, C 60–74, D 40–59, F below 40. A confirmed security flaw caps the score outright — a server with a verified command-injection sink cannot exceed 20 no matter how good the rest of it is — so D and F overwhelmingly indicate a specific, evidenced problem rather than a general lack of polish.
Are month-over-month deltas comparing the same servers?
No. Distribution deltas cover the whole corpus, which grows as new servers are ingested, so part of any shift is composition rather than servers changing. The 'biggest score drops' section is the cohort-stable view: those are specific servers that were re-scanned and scored lower than before.
Why are some listed servers not scored at all?
If a server declares no public repository, or its source cannot be retrieved because it is private, moved, or removed, there is nothing to verify. Those are listed as unverified with no score rather than given an invented number, and they are excluded from every percentage on this page.
What counts as 'no authentication'?
A server that declares no authentication mechanism — no OAuth 2.1, no protected-resource metadata — so nothing gates who may invoke its tools. Separately, among servers whose live endpoints we probed, some declare that auth is required and then serve their full tool surface to an anonymous caller; that is reported as 'auth declared but not enforced'.
Can I use these numbers in my own work?
Yes. The full dataset behind each edition is downloadable as JSON and CSV from the page itself. Cite the Canopii Trust Index and link to the specific monthly edition you used, so the numbers you quoted remain verifiable.

Take the data

Every number on this page is downloadable. Attribution to the Canopii Trust Index is all we ask.

Looking for a specific server? Browse the full directory — every server page shows its complete control checklist with the evidence behind each result.