The State of MCP Security, 2026
Report version 1.3 · Published 2026-08-17, revised 2026-10-15, 2026-10-28 and 2026-11-10 · Scan date 2026-07-26 · Corpus: 2,303 distinct public MCP server configs · Tool: AgentAuditKit (offline, deterministic) · Licence: Apache-2.0 (same as the tool)
Stable identity. Cite this document as The State of MCP Security, 2026, version 1.3. The version changes only when a number in it changes; re-running the pipeline on the same committed corpus reproduces this exact file, so version 1.3 always refers to these numbers. Version 1.3 supersedes version 1.2 (2026-10-28): the corpus and the harness are the same, and a rule added in agent-audit-kit 0.6.34 found a browser-automation server with the user's own browser in reach in six configs. Version 1.2 had superseded 1.1 (2026-10-15) after a rule added in 0.6.28 found a critical flaw in one more config, and 1.1 had superseded 1.0 (2026-08-17) after corrected rules and a corrected scan harness moved the figures. What changed in v1.3 lists every number that moved and why. See How to cite this report.
We statically scanned 2,303 distinct public Model Context Protocol server configurations — the largest snapshot we know of — and measured what fraction fail each security rule family. The headline: more than half of public MCP servers declare a remote endpoint with no authentication. Every number below carries its numerator, denominator, and — where a metric is not computable for the whole corpus — its coverage.
Methods in one paragraph. The corpus was collected 2026-07-26 from two
provenance-tracked sources — 664 GitHub-crawled .mcp.json files
(benchmarks/data/) and 1,639 latest-version servers from the official MCP Registry
(registry.modelcontextprotocol.io) — deduplicated by SHA-256 of the normalised
JSON. Each config was written alone into an empty directory as .mcp.json, scanned
with AgentAuditKit's engine and graded on the same penalty-based A–F scale as
aak score. The scan makes zero network calls and is deterministic: the
same corpus yields a byte-identical results.json across runs and Python hash
seeds. Reproduce the published numbers with make report (offline); refreshing the
corpus itself is a separate, explicitly network step
(python research/state-of-mcp-2026/fetch_registry.py --target 5000). Full detail,
including what a config-only corpus cannot measure, is in Method and
Limitations.
Headline findings
| Finding | Rule | Share of corpus |
|---|---|---|
| Remote MCP server with no authentication | AAK-MCP-001 (critical) |
50.6%, 1,166 of 2,303 |
| Carries at least one critical finding | — | 51.5% (1,185) |
Fetches and executes a package at launch via npx/uvx (supply-chain surface) |
AAK-MCP-005 (medium) |
19.5% (450) |
| OAuth surface with no RFC 9728 Protected-Resource-Metadata discovery | AAK-OAUTH-008 (low) |
18.4% (424) |
| Launches a package with no pinned version | AAK-SUPPLY-001 (medium) |
12.3% (284) |
| Command uses a relative path (PATH-hijack surface) | AAK-MCP-006 (medium) |
7.8% (179) |
| stdio launcher injection: shell interpreter with exec flag / interpolated args | AAK-MCP-STDIO-LAUNCHER-INJECT-001 (high) |
2.1% (49) |
| Deprecated SSE transport (2026-07-28 stateless break) | AAK-TRANSPORT-003 (medium) |
2.0% (47) |
| Secret in the env block (hardcoded credential) | AAK-MCP-003 (high) |
1.3% (30) |
| URL contains an internal network address (SSRF surface) | AAK-MCP-009 (high) |
0.1% (3) |
Auth posture, the 2026-07-28 profile:
- No authentication: 1,166 of 2,303 (50.6%).
- RFC 9728 PRM discovery: 0 of 2,303 (0.0%). This counts configs that reference the discovery document; it does not say whether a server serves it. A live probe of the corpus's public endpoints on 2026-10-11 found that more than half of those that answered DNS do (AAK-MCP-001 measurement).
- Remote configs: 1,648 of 2,303 (71.6%). Of the 424 that embed an inline credential, 424 of 424 (100%) hardcode a static secret rather than reference an env var or use OAuth.
Method
Corpus construction
Two provenance-tracked sources, deduplicated by configuration content (SHA-256 of the normalised JSON):
| Source | Candidates | In corpus |
|---|---|---|
GitHub-crawled .mcp.json files (benchmarks/data/) |
748 | 664 |
Official MCP Registry, latest-version servers (registry.modelcontextprotocol.io) |
1,641 | 1,639 |
| Combined | 2,389 | 2,303 |
85 byte-duplicate configs and 1 unparseable file were removed. Registry servers
were fetched on 2026-07-26 via cursor pagination (40-page crawl), filtered to
isLatest = true / status = active, and each converted to a scannable
.mcp.json-shaped config. A server with a remote endpoint becomes one http or
sse entry for that endpoint, carrying the names of the headers it declares. A
package server launches the release its registry record states:
npx -y <package>@<version> for npm, uvx <package>@<version> for PyPI, and
docker run of the image for OCI. The registry has grown fast — this snapshot is
1,641 distinct latest-version registry servers, up from 710 on 2026-07-19
(2.3×). Provenance for every registry server — name, version, transport, auth
mode, source URL, fetch date — is recorded in corpus/registry-manifest.json,
so each number here is reproducible from the committed manifest without
re-fetching. The manifest is rebuilt offline from the raw registry pages the fetch
cached (fetch_registry.py --from-cache), so a change to the conversion does not
move the snapshot.
Scanning
Each config is written alone into an empty directory as .mcp.json, the name the
crawler saved it under and one that every AgentAuditKit scanner that reads MCP
configs opens, then scanned with AgentAuditKit's engine and scored with the same penalty-based A–F grade the
aak score command and the MCP Security Index use. The scan is offline (zero
network calls) and deterministic: the same corpus yields a byte-identical
results.json on every run, across Python hash seeds — the property that makes a
published number auditable.
Note on what is and isn't config-detectable: rule families that live in a
client .mcp.json — no-auth, launcher injection, relative-path commands, env
secrets, npx/uvx fetch-execute, unpinned launches, CVE version pins,
transport — fire here. Families that require
server source (tool-poisoning AAK-POISON-*, taint AAK-TAINT-*, the
AAK-MCP-STDIO-CMD-INJ-* source detectors) do not fire on a config-only corpus
and are reported as ~0% by construction, not by absence of risk. The MCP CVE
version pins (agent_audit_kit/scanners/mcp_cve_pins_2026_07.py, and the older
pins in supply_chain.py) fire when a config launches a package at an affected
version, or names one whose fixed version it cannot show, such as an unpinned git
launch. A registry config launches the version its record states, so a pin fires
there only on an affected release.
The top-misconfiguration table and the OWASP MCP tally leave out two advisory
posture rules, AAK-MCP-ATTEST-001 (deny-by-default attestation, which fires on
nearly every config) and AAK-MCP-007 (no version pin in args), so neither
figure is inflated. results.json lists them under excluded_advisory_rules, and
its per-rule rule_hit_rate keeps them.
Reproduce
make report # refreshes results.json from the committed manifest, offline + deterministic
# or, explicitly:
python research/state-of-mcp-2026/run_report.py \
--corpus benchmarks/data \
--registry-manifest research/state-of-mcp-2026/corpus/registry-manifest.json \
--out research/state-of-mcp-2026/results.json
Refreshing the corpus itself (the one network step) is separate:
python research/state-of-mcp-2026/fetch_registry.py --target 5000.
At the published run this walk collected 1,641 distinct latest-version servers
(corpus/registry-manifest.json → distinct_latest_servers, fetched_at
2026-07-26). --target 5000 is deliberately larger than that count so the walk
runs to cursor-exhaustion rather than stopping early; the MCP Registry grows over
time, so a later rerun that returns a larger N is registry growth, not a broken
command. (--target is kept identical across the Makefile, the fetcher default, and
this document by tests/test_corpus_target_consistency.py.)
Grade distribution (score calibration)
| Grade | Configs | Share |
|---|---|---|
| A | 787 | 34.2% |
| B | 1,281 | 55.6% |
| C | 96 | 4.2% |
| D | 52 | 2.3% |
| F | 87 | 3.8% |
n = 2,303. The median config scores B. 1,185 (51.5%) carry at least one critical-severity finding — driven almost entirely by the no-auth remote server class. Most public configs are not catastrophic, but the single most common public posture is a remote server anyone can call.
Why this matters now — and the two wedges
The market backdrop is a supply-chain and trust crisis, not a scanner gap:
- Shai-Hulud 2.0 — the self-propagating npm worm that republishes legitimate
packages with a malicious
preinstallscript and uses GitHub infrastructure as C2 — compromised 25,000+ repositories across ~350 maintainers (Zapier, PostHog, Postman), with a Mini Shai-Hulud resurgence in May 2026 hitting 170+ npm packages (Microsoft Security). Our finding that 19.5% of public MCP configs fetch and execute a package at launch throughnpx/uvx, and 12.3% launch one with no pinned version, is that same blast radius, one.mcp.jsonaway. - NSA MCP hardening guidance — the NSA AISC MCP Security Design Considerations CSI (U/OO/6030316-26, 2026-05-20) put MCP posture on government letterhead. It is prose, not a checklist — so the gap is evidence, not awareness.
Against that backdrop, two things a free hosted scanner structurally can't match:
- Offline + deterministic scanning. Zero network calls in the default scan path; the same input always yields the same finding (no model in the loop), so CI diffs, audit re-runs, and regression baselines stay byte-stable. This report is itself the proof: every number regenerates identically from the committed manifest.
- A compliance-evidence crosswalk. Every rule is mapped — rule by rule — to
the NSA MCP Security CSI control it evidences and its OWASP Agentic
Top-10 (2026) slot (also OWASP MCP Top-10 and EU AI Act), emitted as
machine-readable evidence:
agent-audit-kit --emit-coverage --format json(seedocs/coverage.jsonanddocs/crosswalk/nsa-csi-owasp-agentic.md). That turns a finding into audit evidence an assessor can cite.
We are not claiming to be first — public MCP-security surveys and vendor disclosures predate this. What is offline, deterministic, and standards-mapped is the combination.
2026-07-28 stateless-transport exposure
The ratified 2026-07-28 spec removes the initialize handshake and the
Mcp-Session-Id header (stateless core). The measurable proxy in a client config
is deprecated SSE transport: 47 configs (2.0%) still declare it
(AAK-TRANSPORT-003, from a type or transport key or an /sse endpoint) and
will break when the session model goes. streamable-HTTP dominates the fresh
corpus (1,726 server entries) over stdio (1,347) and SSE (62).
Responsible-disclosure posture
Findings here are aggregate and de-identified; no server is named as vulnerable.
The corpus is public registry metadata and public GitHub .mcp.json files. AAK's
90-day disclosure policy governs any per-server
notification.
Limitations
- Config-only for source-based families. Tool-poisoning and taint need server source; their ~0% here is a coverage boundary, not an all-clear (stated above).
- Registry snapshot is a 40-page crawl (1,641 distinct latest); the registry is larger and still growing, so the true corpus is a floor.
- Registry configs are synthesised from the registry record. Each takes the record's first remote that declares headers (or its first remote), or its first package. Header values, package arguments and runtime hints are not carried over, so a server whose real config differs from that shape is measured on the shape.
- Auth posture is inferred from declared config, not a live probe — a server could enforce auth the config doesn't declare (and vice-versa).
- "Benign" is not claimed. This measures declared posture, not exploitability; see the benign-slice false-positive rate for how precise the high-severity findings are.
What changed in v1.3
Version 1.3 rescans the same 2026-07-26 corpus with agent-audit-kit 0.6.34. No config was added or removed, and the harness did not change. One step moved figures:
- (x)
AAK-MCP-BROWSER-ATTACH-001, new in 0.6.34, reports a browser-automation MCP server that has the user's own browser in reach: it attaches over CDP to a browser on the same machine or to the running browser, points--user-data-dirat the real profile, imports cookies, or turns off the renderer sandbox with--no-sandbox. It fires on 6 configs, each one MEDIUM finding, so medium findings rise by 6 and the configs under OWASP's MCP02:2025 rise from 14 (0.6%) to 20 (0.9%). One of the six moves from grade A to B. No other rule added since v1.2 fired on the corpus.
Every figure since v1.0, and the step that moved it; steps (0) to (viii) are v1.1's and (ix) is v1.2's:
| Figure | v1.0 | v1.1 | v1.2 | v1.3 | What moved it |
|---|---|---|---|---|---|
Configs with a remote server and no auth (AAK-MCP-001) |
1,201 (52.1%) | 1,166 (50.6%) | 1,166 (50.6%) | 1,166 (50.6%) | (ii) -34 · (iii) -1 |
| Configs with a critical finding | 1,213 (52.7%) | 1,184 (51.4%) | 1,185 (51.5%) | 1,185 (51.5%) | (ii) -34 · (iii) -1 · (v) +6 · (ix) +1 |
| Configs with a high finding | 165 (7.2%) | 110 (4.8%) | 110 (4.8%) | 110 (4.8%) | (ii) -37 · (iv) -36 · (v) +18 |
Configs launching a package with no pinned version (AAK-SUPPLY-001) |
0 (0.0%) | 284 (12.3%) | 284 (12.3%) | 284 (12.3%) | (v) +444 · (vi) -160 |
Configs with an env secret, AAK-MCP-003 |
70 (3.0%) | 30 (1.3%) | 30 (1.3%) | 30 (1.3%) | (iv) -40 |
Configs with an env secret, AAK-SECRET-007 |
70 (3.0%) | 0 (0.0%) | 0 (0.0%) | 0 (0.0%) | (v) -70 |
Configs with an internal-network URL (AAK-MCP-009) |
44 (1.9%) | 3 (0.1%) | 3 (0.1%) | 3 (0.1%) | (ii) -41 |
Configs declaring SSE (AAK-TRANSPORT-003) |
36 (1.6%) | 47 (2.0%) | 47 (2.0%) | 47 (2.0%) | (viii) +11 |
| Grade A | 637 (27.7%) | 789 (34.3%) | 788 (34.2%) | 787 (34.2%) | (0) -1 · (ii) +21 · (iv) +2 · (v) -28 · (vi) +160 · (viii) -2 · (ix) -1 · (x) -1 |
| Grade B | 1,421 (61.7%) | 1,279 (55.5%) | 1,280 (55.6%) | 1,281 (55.6%) | (ii) +5 · (iii) +1 · (iv) +10 · (v) +8 · (vi) -160 · (viii) -6 · (ix) +1 · (x) +1 |
| Grade C | 111 (4.8%) | 96 (4.2%) | 96 (4.2%) | 96 (4.2%) | (ii) -13 · (iv) -10 · (viii) +8 |
| Grade D | 49 (2.1%) | 52 (2.3%) | 52 (2.3%) | 52 (2.3%) | (ii) -8 · (v) +11 |
| Grade F | 85 (3.7%) | 87 (3.8%) | 87 (3.8%) | 87 (3.8%) | (0) +1 · (ii) -5 · (iii) -1 · (iv) -2 · (v) +9 |
| Critical findings | 1,335 | 1,292 | 1,293 | 1,293 | (ii) -46 · (iii) -3 · (v) +6 · (ix) +1 |
| High findings | 319 | 208 | 208 | 208 | (ii) -49 · (iv) -85 · (v) +23 |
| Medium findings | 4,432 | 4,930 | 4,930 | 4,936 | (v) +645 · (vi) -160 · (viii) +13 · (x) +6 |
| Low findings | 1,224 | 652 | 652 | 652 | (v) -572 |
| Configs with a secret-exposure finding | 70 (3.0%) | 4 (0.2%) | 4 (0.2%) | 4 (0.2%) | (v) -66 |
| Configs with a transport-security finding | 42 (1.8%) | 53 (2.3%) | 53 (2.3%) | 53 (2.3%) | (viii) +11 |
| Configs with a supply-chain finding | 0 (0.0%) | 287 (12.5%) | 288 (12.5%) | 288 (12.5%) | (v) +447 · (vi) -160 · (ix) +1 |
| Configs with a legal-compliance finding | 0 (0.0%) | 8 (0.3%) | 8 (0.3%) | 8 (0.3%) | (v) +8 |
What changed in v1.2
Version 1.2 rescans the same 2026-07-26 corpus with agent-audit-kit 0.6.28. No config was added or removed, and the harness did not change. One step moved figures:
- (ix)
AAK-RUNTIME-ADV-001, new in 0.6.28, reads the runtime an MCP server is launched with. One config starts its server from an nvm install of Node.js 22.22.3, below 22.23.4, the 22 line's fix for CVE-2026-111497 (an out-of-bounds write in the Brotli decoder Node.js bundles, in CISA's Known Exploited Vulnerabilities catalog since 2026-10-27). It is that config's first critical finding, so it moves from grade A to B. No other rule added since v1.1 fired on the corpus.
The table of every figure since v1.0, and the step that moved it, is under What changed in v1.3.
What changed in v1.1
Version 1.1 rescans the same 2026-07-26 corpus. No config was added or removed. Every number that moved is below, with the step that moved it. The steps were applied one at a time to the v1.0 code, in the order listed, and every figure was recorded after each.
- (0) Version 1.0's grade table disagreed with its own
results.json: it kept an earlier run's A and F counts, 637 and 85, where the data said 636 and 86. - (i) Rule changes merged after the code that produced v1.0, through #991. They changed no config's findings on this corpus; they moved only the OWASP breakdown (#971, below).
- (ii)
AAK-MCP-001andAAK-MCP-009no longer report a loopback server (localhost, 127.0.0.0/8, ::1, 0.0.0.0), which is the developer's own process. - (iii)
AAK-MCP-001recognises more API-key header names: any name ending in_API_KEYorApiKey, and a-Subscription-Keysuffix. - (iv)
AAK-MCP-003reports only literals that can be credentials, not settings such as*_TOKEN_NAMEor short demo values. - (v) The harness wrote each config as
config.mcp.json. Seven scanners read MCP configs by name and never open that one, among them the supply-chain scanner and the CVE pin table, so up to 128 rules were never measured on the corpus. AndAAK-SECRET-007, which stands down on.mcp.jsonbecauseAAK-MCP-003reports the same env literal there, counted every env secret a second time. Each config is now scanned as.mcp.json. - (vi) The registry conversion dropped the version each record states, so its 161 npm and PyPI configs were unpinned launches by construction. They now launch the stated release.
- (vii) The OWASP MCP tally leaves out the two advisory rules, as the top-misconfiguration table already did.
- (viii)
AAK-TRANSPORT-003reads thetypekey, where Claude Code and VS Code declare SSE, as well astransport, and reports each server once.
The table under What changed in v1.3 gives every figure at v1.0, v1.1, v1.2 and v1.3, with the step that moved it.
The new findings in (v) are AAK-SUPPLY-001, eight CVE version pins (the Serena
pin, on 19 configs that launch Serena from its git repository or a local checkout
at no stated version, enters the top ten), and the EU AI Act Art. 50 transparency
rules on 8 configs. AAK-MCP-007 stands
down wherever AAK-SUPPLY-001 reports the same launch, and that is the whole
low-severity drop. In (vi) the 160 registry configs that launched an unpinned
package lose that finding, and each moves from B to A.
Statements v1.0 made that the corrected scan does not support:
- "19.5% of public MCP servers fetch-and-execute unpinned remote packages." The
19.5% is
AAK-MCP-005, which counts everynpx/uvxlaunch, pinned or not. Unpinned launches areAAK-SUPPLY-001, 12.3%, which v1.0 never measured. - "Not a single public server serves the discovery document." The figure counts configs, which cannot serve anything; the 2026-10-11 probe found that more than half of the public endpoints that answered DNS serve it.
- "Pinned-CVE hits are rare because registry configs use unpinned
npx -y." The pins never ran on the corpus in v1.0 (step v), and the registry configs were unpinned only because the conversion dropped their versions (step vi).
Where order matters. The combined effect does not depend on the order of the steps; how it splits between two of them can:
- (iv) and (v) remove the same double-counted env literal from two sides:
AAK-MCP-003at high severity in (iv),AAK-SECRET-007at medium in (v). In this order the OWASP MCP01 tally moves only in (v) (-37); the other way round it moves in (iv) (-38) and (v) adds 1. Grades and the count of configs with a high finding split differently too. - (v) and (vi): the versions change only what
AAK-SUPPLY-001reports, and that rule reads the corpus only from (v) on. In this order (v) addsAAK-SUPPLY-001to 444 configs and (vi) removes it from 160. The other way round, (vi) first removesAAK-MCP-007(low) from 160 configs and (v) then addsAAK-SUPPLY-001to - The grade moves are the same either way.
- (ii) and (v): 2 configs lose their only critical finding in (ii) and gain a critical CVE pin in (v), so (ii) moves the critical count by -34 in this order and -32 after (v). (iii) and (vii) each shift one or two configs the same way.
- (viii) moves the same figures whether it is applied first or last.
The OWASP MCP breakdown changed meaning as well as value. v1.0 tallied AAK's
own MCPnn labels, which named different risks from OWASP's list; #971 remapped
every rule to the risk OWASP names (step i). Under OWASP's ids, with the advisory
rules left out (step vii), the most common risk is MCP07, Insufficient
Authentication & Authorization, at 69.3% (1,595). v1.0's table put its own
MCP07 label at 99.8%. The full v1.3 table is in PREVALENCE.md.
How to cite this report
Cite the report, not the repository, when you are citing a number from it — the repository moves and the report does not. Version 1.3 refers to the 2026-07-26 scan of 2,303 configs described above.
Plain text
Jain, S. (2026). The State of MCP Security, 2026 (Version 1.3). AgentAuditKit. https://github.com/sattyamjjain/agent-audit-kit/blob/main/research/state-of-mcp-2026/REPORT.md
BibTeX
@techreport{jain2026mcpsecurity,
author = {Jain, Sattyam},
title = {The State of MCP Security, 2026},
institution = {AgentAuditKit},
year = {2026},
month = {11},
type = {Technical Report},
number = {Version 1.3},
url = {https://github.com/sattyamjjain/agent-audit-kit/blob/main/research/state-of-mcp-2026/REPORT.md},
note = {Corpus: 2,303 distinct public MCP server configurations, collected 2026-07-26. Offline, deterministic scan; reproducible with \texttt{make report}.}
}
Repository-level metadata is in CITATION.cff at the repo
root, which GitHub renders as a "Cite this repository" widget. It points at this same
identity, so the two cannot disagree about the report version.
Citing a specific number. Every figure above carries its numerator and
denominator; quote both. For example: "0 of 2,303 public MCP configs referenced
RFC 9728 Protected Resource Metadata discovery (Jain 2026, v1.3)." If you need the
raw data, results.json beside this file is the machine-readable output, and
corpus/registry-manifest.json records per-server provenance.