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, andslackwere 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 underoriginal_scoring.
Revision 2026-10-16 (ruleset 1.2.0) - the destructive-tool findings below were overstated.
MCP_DESTRUCTIVE_UNCONSTRAINEDproduced 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 MCPreadOnlyHint/destructiveHintannotations. The static extractor also misread//comments containing an apostrophe, which hid tools and arguments. Ruleset 1.2.0 fixes all three:
readOnlyHint: truemeans the tool is not destructive.- A destructive tool that declares
destructiveHint: trueis 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; firecrawlfirecrawl_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-commanderDesktopCommanderMCP f508e52(2026-10-14)25 HIGH, 8 HIGH findings MODERATE, no HIGH finding. Every state-changing tool declares destructiveHint: trueand every read toolreadOnlyHint: true; 9 MEDIUM "host-side approval only" findings remain.firecrawlfirecrawl-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).filesystemmodelcontextprotocol/servers bbb701e(2026-10-11), annotations read from source14 HIGH ( write_file)No HIGH finding. write_file,edit_fileandmove_filedeclaredestructiveHint: 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: falsebut 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_agentandfirecrawl_monitor_runstart credit-consuming jobs. The rest of this report (other servers, headline numbers) still reflects ruleset 1.1.0 until a full re-scan.
- 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.
| Rule | Occurrences |
|---|---|
MCP_UNBOUNDED_PARAMETER |
242 |
MCP_DESTRUCTIVE_UNCONSTRAINED |
25 |
MCP_MUTATING_NO_REQUIRED |
3 |
| 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 |
desktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - tooledit_blockdesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolget_promptsdesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolinteract_with_processdesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toollist_sessionsdesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolstart_processdesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolstart_searchdesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolwrite_filedesktop-commanderMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolwrite_pdfeverythingMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolsimulate-research-queryfilesystemMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolwrite_filefirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_agentfirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_credit_usagefirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_developer_searchfirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_monitor_createfirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_monitor_deletefirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_monitor_runfirecrawlMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfirecrawl_searchgithubMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolmerge_pull_requestmemoryMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - tooldelete_entitiesmemoryMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - tooldelete_observationsmemoryMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - tooldelete_relationsplaywrightMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolbrowser_dragplaywrightMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolbrowser_dropplaywrightMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolbrowser_run_code_unsafeserver-fetchMCP_DESTRUCTIVE_UNCONSTRAINED(HIGH): Destructive tool exposes no authorization surface - toolfetch
- Live probes launch each published package exactly as a client would (
npx -y <package>/uvx <package>) and readtools/listoff the wire after a full MCP handshake. Nothing is executed on the target beyond protocol-level requests. - 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. - 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/. - 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.
- 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.
- 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.
- This scan covers a curated set of high-adoption servers, not the full ecosystem registry; percentages are for this cohort.
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.