State of MCP Security 2026 — Prevalence & Score Calibration
We scanned 2,303 distinct public MCP server configs with an offline static scanner. More than 1 in 2 declares a remote server with no authentication, more than half ship a critical-severity flaw, the median config scores a B, and the top 10% score an A.
Reproducible data report (empirical companion to the State-of-MCP-Security report, issue #23). Every "measured" number below comes from running AgentAuditKit (fully offline, deterministic, Apache-2.0; the exact ruleset is the committed
rules.jsonbundle at this repo's HEAD) over a content-deduplicated corpus of public MCP configs, and is the figure of report version 1.3. Raw aggregate:results.json. External figures are attributed to their sources, not measured by us.On the corpus. Two provenance-tracked sources — a GitHub Code Search crawl of public MCP config files and the official MCP Registry's latest-version servers — deduplicated by file content: 2,303 distinct configs. This is the same corpus
REPORT.mddocuments; the split and provenance are in the Methodology section below. We report the real N — 2,303 — not a round target.
Three headline numbers
- 50.6% (1,166/2,303) of public MCP configs declare a remote server with no
authentication — a critical-severity posture (
AAK-MCP-001) and the single most common finding in the corpus. It corroborates Knostic's 119-of-119 unauthenticated tool-listing finding from the deployment side. - 51.5% (1,185/2,303) contain at least one critical-severity finding — driven almost entirely by that no-auth remote-server class.
- The median public MCP config scores a B; the top 10% score an A. (This is the empirical score-calibration anchor issue #23 asked for. The roadmap guessed "median C-, top 10% B+" — the real distribution skews higher, because most configs carry one dominant critical posture rather than a long tail of findings.)
Score distribution (A–F) — the calibration
| Grade | Configs | Share | Cumulative |
|---|---|---|---|
| A | 787 | 34.2% | 34.2% |
| B | 1,281 | 55.6% | 89.8% |
| C | 96 | 4.2% | 94.0% |
| D | 52 | 2.3% | 96.2% |
| F | 87 | 3.8% | 100% |
- Median grade: B, the grade of config 1,152 of 2,303 ranked best first (cumulative A = 787, A+B = 2,068).
- Top 10% (the best 230 configs): none below A (A alone holds 787).
- 6.0% (139) land at D or F.
The grade is AAK's penalty-based score (start at 100; deduct 20/CRITICAL,
10/HIGH, 5/MEDIUM, 2/LOW), identical to aak score. **Calibration takeaway for
23:** the current penalties produce a sane spread — a clear A/B mode with a real
D/F tail — and the median lands at B, not the "C-" the roadmap assumed. What moved the story since the earlier crawl-only corpus is not the penalty weights but the corpus: adding the MCP Registry surfaced how dominant the no-auth remote-server posture is. No penalty re-weighting is done in this report (that's a follow-up).
Severity + category prevalence
All findings: 1,293 critical · 208 high · 4,936 medium · 652 low. Per-config: 51.5% have ≥1 critical; 4.8% have ≥1 high.
| AAK category | Configs with ≥1 finding | Share |
|---|---|---|
| MCP Configuration | 2,300 | 99.9% |
| Supply Chain | 288 | 12.5% |
| Transport Security | 53 | 2.3% |
| Legal Compliance | 8 | 0.3% |
| Secret Exposure | 4 | 0.2% |
| Composition | 1 | 0.0% |
| Tool Poisoning | 1 | 0.0% |
OWASP MCP Top 10 — configs tripping each risk
Each rule carries the OWASP MCP Top 10 ids (v0.1, the 2025 edition) of the risks it detects, under the titles OWASP gives them. A config trips a risk when one of its findings carries the id; the two advisory-posture rules below are left out, as in the top-ten table. Before report version 1.1 this table used AAK's own labels for the ids, in a different order, so the two do not compare row for row.
| OWASP MCP Top 10 | Risk | Configs | Share |
|---|---|---|---|
| MCP07:2025 | Insufficient Authentication & Authorization | 1,595 | 69.3% |
| MCP01:2025 | Token Mismanagement & Secret Exposure | 455 | 19.8% |
| MCP04:2025 | Software Supply Chain Attacks & Dependency Tampering | 453 | 19.7% |
| MCP05:2025 | Command Injection & Execution | 213 | 9.2% |
| MCP02:2025 | Privilege Escalation via Scope Creep | 20 | 0.9% |
| MCP09:2025 | Shadow MCP Servers | 3 | 0.1% |
| MCP10:2025 | Context Injection & Over-Sharing | 2 | 0.1% |
| MCP03:2025 | Tool Poisoning | 1 | 0.0% |
Top 10 most-common findings (advisory-posture rules excluded)
| # | Rule | Finding | Severity | Configs | Share |
|---|---|---|---|---|---|
| 1 | AAK-MCP-001 |
Remote MCP server without authentication | CRITICAL | 1,166 | 50.6% |
| 2 | AAK-MCP-005 |
MCP server uses npx/uvx to fetch and execute remote packages | MEDIUM | 450 | 19.5% |
| 3 | AAK-OAUTH-008 |
MCP OAuth surface with no RFC 9728 Protected Resource Metadata discovery | LOW | 424 | 18.4% |
| 4 | AAK-SUPPLY-001 |
MCP server package not pinned to exact version | MEDIUM | 284 | 12.3% |
| 5 | AAK-MCP-006 |
MCP server command uses relative path | MEDIUM | 179 | 7.8% |
| 6 | AAK-MCP-STDIO-LAUNCHER-INJECT-001 |
MCP stdio server launches a shell-style interpreter with an exec flag or interpolated args | HIGH | 49 | 2.1% |
| 7 | AAK-TRANSPORT-003 |
Deprecated SSE transport in use | MEDIUM | 47 | 2.0% |
| 8 | AAK-MCP-003 |
MCP server environment exposes secrets | HIGH | 30 | 1.3% |
| 9 | AAK-MCP-SERENA-CVE-2026-49471-001 |
Serena MCP toolkit below 1.5.2 (unauthenticated dashboard → DNS-rebinding RCE) | HIGH | 19 | 0.8% |
| 10 | AAK-MCP-004 |
Excessive number of MCP servers declared | HIGH | 12 | 0.5% |
Two advisory-posture rules are excluded so the story isn't inflated (tracked
in results.json under excluded_advisory_rules): AAK-MCP-ATTEST-001
(deny-by-default attestation — fires on ~every server) and AAK-MCP-007 (no
version pin in args, LOW). The two fixes that matter, in order: authenticate
remote servers (50.6% declare one with no auth) and pin what you launch
(12.3% launch a package with no pinned version, and 19.5% fetch and execute one
through npx/uvx at launch).
Methodology
- Corpus. MCP config files from two provenance-tracked sources — a GitHub Code
Search crawl across five filename queries (
.mcp.json,mcp.json,claude_desktop_config.json,cline_mcp_settings.json,mcp_settings.json) and the official MCP Registry's latest-version servers — deduplicated by repo+path at crawl time and again by file content at scan time.
In-corpus split: 664 crawled + 1,639 registry = 2,303 distinct, from 2,389 candidates (85 byte-identical duplicates + 1 unparseable removed).
- Registry configs. Each registry server becomes one scannable config: its
remote endpoint with the names of the headers it declares, or its package
launched at the version the registry record states (npx -y <package>@<version>,
uvx <package>@<version>).
- Scanner. AgentAuditKit, offline and deterministic — no cloud, LLM, or
telemetry. Each config is written alone as .mcp.json, the name every scanner
reads, and scanned via the same run_scan + compute_score the CLI and MCP
Security Index use. Deterministic: same corpus → byte-identical results.json.
- Public metadata only. We scan committed configuration files and public
registry metadata. No live server probing, no exploitation, no credential use.
Rate-limited crawl, respects GitHub API limits.
Reproducibility
- Ruleset: the committed, Sigstore-signable
rules.jsonbundle at this repo's HEAD. - Corpus: regenerate
results.jsonfrom the committed manifest, offline and deterministically. We deliberately do not publish a per-server list mapping repos to grades/findings — that would de-anonymize specific servers, against the repo's 90-day coordinated-disclosure policy. The report is aggregate-only.
git clone https://github.com/sattyamjjain/agent-audit-kit && cd agent-audit-kit
pip install -e .
make report # refreshes results.json from the committed manifest, offline + deterministic
# 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 the registry 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 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.
Scan your own in 30s, fully offline: pip install agent-audit-kit && agent-audit-kit scan .
Limitations (read before quoting)
- Public-metadata static scan, not runtime. A matched pattern is not proof of exploitability in context; a clean scan is not proof of safety. Auth posture is inferred from declared config, not a live probe.
- N of unreachable servers. Downloads fail for private/deleted repos and rate-limited fetches; those are dropped, so 2,303 is "distinct public configs we could fetch and parse," not "all MCP servers."
- Sample skews to discoverable public repos + registry latest-version entries (Code Search ranks by relevance + stars; the registry crawl is a 40-page snapshot). Not a uniform random sample of all registered servers.
- External figures are as-reported by the cited third parties (Knostic 1,862 exposed / 119-of-119 unauthenticated; the 2,614-server 82%-path-traversal survey). Only the 2,303-config scan is ours.
- No CVEs/advisories were filed for this report. A static match on a committed config is not live-exploitation proof, and disclosing a specific server is a deliberate, per-maintainer coordinated-disclosure decision — not something to automate off an aggregate scan. Genuine exposures surface through the MCP Security Index under the 90-day policy.
Distribution
Ready-to-post launch copy lives in one place —
docs/DISTRIBUTION-CHECKLIST.md — and every
quantitative token there is substituted from results.json. This report is the
data; the checklist is the copy. Nothing here or there auto-posts.