Skip to content

Latest commit

 

History

History
129 lines (103 loc) · 13.9 KB

MCP_ECOSYSTEM_SCAN.md

File metadata and controls

129 lines (103 loc) · 13.9 KB

MCP Ecosystem Security Scan

Generated: 2026-10-03 12:52 UTC
Scanner: findings from LIFE FORGE mcpsec ruleset 1.0.0; scores and bands re-computed with mcpsec 1.1.0 (see revision note) Method: live protocol probes (stdio, definitions read off the wire) for servers launchable without credentials; static source extraction for popular credential-gated servers. The scanner observes definitions only - it never executes tools.

Revision 2026-10-10 - read first. This report was first published (2026-10-03) with 40% of servers in the CRITICAL band. That figure was wrong. The 1.0.0 scorer banded servers by raw finding volume, so a server with many MEDIUM "unbounded parameter" findings and no CRITICAL finding was labelled CRITICAL; the scan contained zero CRITICAL-severity findings. Bands are now anchored on the most severe finding (mcpsec 1.1.0), and the stored findings were re-scored without re-collecting them. Two further 1.0.0 defects mean the findings themselves need a fresh scan: (1) static extraction dropped every parameter of tools declared as JSON-schema constants, so brave-search, google-maps, and slack were never actually scanned (their 0-finding results are invalid); (2) live scans collapsed per-parameter findings, undercounting MEDIUM findings on live-probed servers. Original scores are preserved in each per-server JSON under original_scoring.

Revision 2026-10-16 (ruleset 1.2.0) - the destructive-tool findings below were overstated. MCP_DESTRUCTIVE_UNCONSTRAINED produced false positives. It matched verbs anywhere in a description, including adjectives ("merged pull requests", "purchased credits") and usage notes ("do not send both fields"), and it ignored the MCP readOnlyHint / destructiveHint annotations. The static extractor also misread // comments containing an apostrophe, which hid tools and arguments. Ruleset 1.2.0 fixes all three:

  • readOnlyHint: true means the tool is not destructive.
  • A destructive tool that declares destructiveHint: true is MEDIUM, not HIGH: a host can ask the user first, and what remains is that the server enforces no limit of its own.
  • A read-named tool is judged by its name, and only a description's first sentence is matched.

In the per-server findings below, these HIGH flags are false positives: desktop-commander get_prompts, list_sessions, start_search; firecrawl firecrawl_search, firecrawl_developer_search, firecrawl_credit_usage. The three servers with open disclosures were re-checked against current source (static extraction; no server was launched):

Server Source checked Tools In this report (1.1.0) Ruleset 1.2.0
desktop-commander DesktopCommanderMCP f508e52 (2026-10-14) 25 HIGH, 8 HIGH findings MODERATE, no HIGH finding. Every state-changing tool declares destructiveHint: true and every read tool readOnlyHint: true; 9 MEDIUM "host-side approval only" findings remain.
firecrawl firecrawl-mcp-server e895e94 (2026-10-15) 28 HIGH, 7 HIGH findings HIGH, 3 HIGH findings: firecrawl_interact, firecrawl_agent, firecrawl_monitor_run (see below). 61 parameters without bounds (MEDIUM).
filesystem modelcontextprotocol/servers bbb701e (2026-10-11), annotations read from source 14 HIGH (write_file) No HIGH finding. write_file, edit_file and move_file declare destructiveHint: true (MEDIUM under 1.2.0); not fully re-scored, because the static extractor does not parse this server's registration pattern.

The firecrawl HIGH findings are tools that declare destructiveHint: false but start work or act outside the server. firecrawl_interact's own description says form submissions "can create persistent external side effects", so a client that trusts the hint will not ask the user first. firecrawl_agent and firecrawl_monitor_run start credit-consuming jobs. The rest of this report (other servers, headline numbers) still reflects ruleset 1.1.0 until a full re-scan.


Headline Numbers

  • 15 of 16 targeted servers scanned (1 not launchable without credentials, 0 scan errors)
  • 167 tool definitions analyzed
  • 80.0% of scanned servers carry at least one finding (a lower bound: three static extractions were invalid, see revision note)
  • 0.0% of scanned servers score in the CRITICAL band (none)
  • Findings by severity: 0 CRITICAL / 25 HIGH / 245 MEDIUM / 0 LOW

Context: an industry scan reported 33% of MCP servers with critical vulnerabilities (Practical DevSecOps, 2026). This scan runs an independent, reproducible ruleset and publishes every raw report.


Findings by Rule

Rule Occurrences
MCP_UNBOUNDED_PARAMETER 242
MCP_DESTRUCTIVE_UNCONSTRAINED 25
MCP_MUTATING_NO_REQUIRED 3

Per-Server Results

Server Source Method Tools Score Band C/H/M/L
brave-search modelcontextprotocol/servers-archived static-source 2 0 LOW 0/0/0/0
context7 @upstash/context7-mcp live-stdio 2 20 MODERATE 0/0/2/0
desktop-commander @wonderwhy-er/desktop-commander live-stdio 26 57 HIGH 0/8/27/0
everything @modelcontextprotocol/server-everything live-stdio 13 50 HIGH 0/1/6/0
filesystem @modelcontextprotocol/server-filesystem live-stdio 14 50 HIGH 0/1/14/0
firecrawl firecrawl/firecrawl-mcp-server static-source 19 59 HIGH 0/7/58/0
github modelcontextprotocol/servers-archived static-source 26 50 HIGH 0/1/93/0
google-maps modelcontextprotocol/servers-archived static-source 7 0 LOW 0/0/0/0
memory @modelcontextprotocol/server-memory live-stdio 9 50 HIGH 0/3/1/0
playwright @playwright/mcp live-stdio 25 50 HIGH 0/3/24/0
sequential-thinking @modelcontextprotocol/server-sequential-thinking live-stdio 1 20 MODERATE 0/0/1/0
server-fetch mcp-server-fetch (PyPI) live-stdio 1 50 HIGH 0/1/1/0
server-git mcp-server-git (PyPI) live-stdio 12 20 MODERATE 0/0/16/0
server-sqlite mcp-server-sqlite (PyPI) live-stdio - - not launchable -
server-time mcp-server-time (PyPI) live-stdio 2 20 MODERATE 0/0/2/0
slack modelcontextprotocol/servers-archived static-source 8 0 LOW 0/0/0/0

Notable Findings

  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool edit_block
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool get_prompts
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool interact_with_process
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool list_sessions
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool start_process
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool start_search
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool write_file
  • desktop-commander MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool write_pdf
  • everything MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool simulate-research-query
  • filesystem MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool write_file
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_agent
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_credit_usage
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_developer_search
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_monitor_create
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_monitor_delete
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_monitor_run
  • firecrawl MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool firecrawl_search
  • github MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool merge_pull_request
  • memory MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool delete_entities
  • memory MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool delete_observations
  • memory MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool delete_relations
  • playwright MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool browser_drag
  • playwright MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool browser_drop
  • playwright MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool browser_run_code_unsafe
  • server-fetch MCP_DESTRUCTIVE_UNCONSTRAINED (HIGH): Destructive tool exposes no authorization surface - tool fetch

Methodology & Limitations

  1. Live probes launch each published package exactly as a client would (npx -y <package> / uvx <package>) and read tools/list off the wire after a full MCP handshake. Nothing is executed on the target beyond protocol-level requests.
  2. Static extraction recovers definitions from source with known registration patterns (server.tool, registerTool, FastMCP @mcp.tool); extraction counters are recorded per server, and servers where no pattern matched are reported as such rather than silently dropped.
  3. Findings are deterministic: re-running the scan against the same server versions reproduces every result. Per-server raw reports ship alongside this document in results/mcp_ecosystem_scan/reports/.
  4. Severity reflects attack surface present in tool definitions, not a proven exploit chain: a CRITICAL here means an attacker-controlled channel or an unconstrained destructive capability exists in the definition itself.
  5. The destructive-tool heuristic matches verbs in tool names and descriptions. It deliberately over-approximates: a read tool whose description mentions execution can be flagged HIGH. Every finding lists its evidence; maintainers reviewing their own tools should read the rule as 'this capability exists without an authorization surface', and open an issue if a flag is wrong.
  6. Live probing executes the server's startup code. The scanner never calls tools on a target, but launching a published package runs whatever its startup performs - observed directly during this scan, when Desktop Commander's first-run flow opened its onboarding page in the analyst's browser, attributing the visit via utm_source=lifeforge-mcpsec read from the MCP initialize handshake. Treat probe targets as untrusted and run them sandboxed.
  7. This scan covers a curated set of high-adoption servers, not the full ecosystem registry; percentages are for this cohort.

Maintainer Engagement & Disclosure Status

In accordance with responsible disclosure principles, findings were shared directly with maintainers prior to broad publicity. As of 2026-10-16, maintainer triage status is recorded below. On 2026-10-16 a correction was posted on each open issue (see the revision note at the top):

Target Repository & Issue Status Outcome / Maintainer Response
Microsoft microsoft/playwright-mcp#1784 Closed (Completed) Reviewed and closed as completed by Playwright Lead Architect Pavel Feldman (@pavelfeldman).
Upstash upstash/context7#3312 Closed (Resolved) Acknowledged with thanks by core collaborator Enes Gules (@enesgules). Proactive correction of the description false positive validated; confirmed that upstream Context7 API validates and bounds inputs prior to execution.
MCP reference servers modelcontextprotocol/servers#4958 Open (Discussion) A contributor review showed write_file already declares destructiveHint: true and is confined by validatePath, and that a confirm argument is self-attestation. Correction posted 2026-10-16: no HIGH finding under ruleset 1.2.0; the confirm suggestion withdrawn; only a documentation question remains.
Desktop Commander wonderwhy-er/DesktopCommanderMCP#808 Open Correction posted 2026-10-16: the HIGH findings were false positives (annotations ignored); no HIGH finding under ruleset 1.2.0. Awaiting maintainer response.
Firecrawl firecrawl/firecrawl-mcp-server#481 Open Correction posted 2026-10-16: three false positives and the parameter count withdrawn. One point stands: firecrawl_interact declares destructiveHint: false although it can submit forms. Awaiting maintainer response.

Scan performed with LIFE FORGE (lifeforge mcp-scan), MIT License. Run it on your own servers before someone else does.