Repository navigation
Releases
v0.23.0
Added
- MCP007: follows paths through
os.path.join,/and tool-argument
reads. A write path built asos.path.join(base, p)orbase / p, or read
from a parameter dict (arguments.get("path"),arguments["path"], and items
of a list read that way in aforloop) now counts as caller-controlled. This
catches CVE-2026-44336 (PraisonAI 4.6.33rules_create) and CVE-2026-84201
(appium-mcp-serverwrite_file/write_files_batch), which 0.22.0 missed. - MCP007: path-builder helpers. A function that returns
os.path.join(base, p)orbase / pof a parameter with no containment
check is a medium finding. This catches CVE-2026-40576 (excel-mcp-server
0.1.7get_excel_path); 0.1.8 stays clean. - MCP007:
x.save(path)(openpyxl, PIL) is a write sink when the path is
joined in the same function.
Fixed
- MCP007: pytest
tmp_path/tmpdirparameters are not caller input.
Verified on real code: catches CVE-2026-40576 (excel-mcp-server 0.1.7), CVE-2026-44336 (PraisonAI 4.6.33) and CVE-2026-84201 (appium-mcp-server 0.1.61), all missed by 0.22.0; patched versions stay clean. SSRF is still out of scope for this rule.
v0.22.0: MCP007 catches the CVE-2026-27825 write shape
What changed
MCP007 now catches writes to a caller-controlled path. A Python function that writes (open(..., "w"/"wb"/"a"/"x"), write_text/write_bytes, shutil.copy*/move, os.rename/replace) to one of its own parameters, or to Path(p)/os.path.abspath(p) of one, with no base-directory check is now a high finding.
This is the shape of CVE-2026-27825 (CVSS 9.1) in sooperset/mcp-atlassian 0.16.0. mcpscan 0.21.0 missed it because nothing on the open() line was interpolated.
| mcp-atlassian | 0.21.0 MCP007 | 0.22.0 MCP007 |
|---|---|---|
| v0.16.0 (vulnerable) | 2 false positives (window.open in HTML strings) |
confluence/attachments.py:198, jira/attachments.py:50 |
| v0.17.0 (patched) | same 2 false positives | none |
Fixed: MCP007 no longer flags browser window.open( inside Python HTML strings.
Limit: a path passed through several functions before the write is still not tracked. A clean scan is not a clean review.
pip install -U mcpscan-cli
v0.21.0
Added
-
MCP024-MCP030: server-side MCP authorization rules. Before this
release, mcpscan's auth-adjacent rules looked at how a client connects to a
server, not at how a server validates the tokens it receives. A deliberately
naive MCP server that fails 8 authorization tests got a clean report from
0.20.0. Newmcpscan/rules/server_auth.py, AST-based (no regex):- MCP024 (high, MCP07):
jwt.decode()with signature verification
disabled (verify_signature: False, legacyverify=False, or"none"in
algorithms). - MCP025 (high, MCP07):
jwt.decode()with noaudience=or with
verify_aud: False. - MCP026 (medium, MCP07):
jwt.decode()with noissuer=or with
verify_iss: False. - MCP027 (low-medium, MCP07): hand-built 401s while nothing in the
project emits aWWW-Authenticate: Bearer resource_metadata=challenge. - MCP028 (medium, MCP07): hand-built bearer auth while no file routes
/.well-known/oauth-protected-resourceand no SDK helper serves it. - MCP029 (high, MCP01): token passthrough. The inbound
Authorization
header reaches an outboundAuthorizationheader, tracked by a small
intra-module taint analysis across returns, aliases and helper
parameters. - MCP030 (medium, MCP01):
access_tokenread from a query object.
Scope gate for precision: the rules only look at Python files that import
mcp/fastmcpor dispatchtools/call/tools/list. When the scan
contains such a server, they also cover modules that read the inbound
Authorizationheader. Test files are skipped. PyJWT and python-jose are
covered; authlib is not. On the naive/hardened experiment servers, 6 of 8
failures are now caught (was 0), with 0 findings on the hardened server.
The two remaining failures, per-tool scope checks and anonymous
tools/list, are runtime-only by nature. Also run against fastmcp, themcp
SDK, litellm, hermes-agent, paperclip and four smaller MCP projects: 0
new-rule findings after fixing two false-positive sources found in
litellm (outbound-request headers mistaken for inbound, and session-cookie
JWT decodes).
43 new tests (tests/test_server_auth.py), plus a vulnerable/clean
auth_server.pyfixture pair. - MCP024 (high, MCP07):
-
README: MCP023 was missing from the rules table; added together with
MCP024-MCP030 and the OWASP mapping.
Fixed
- Inline suppression now honors every
mcpscan: ignore[...]marker on a line,
not just the first.# mcpscan: ignore[MCP002] mcpscan: ignore[MCP023]used
to silence only MCP002. - Dogfood scan is clean again: MCP023 was flagging the attack sample strings in
tests/test_rule_mcp015.py, which had kept CI red since the rule landed.
v0.20.0 - Runtime Security Guards & Provenance Tagging
What's Changed
- feat(runtime): add live tool guards, provenance tagging, and schema drift ledger by @glatinone in #1
New Contributors
- @glatinone made their first contribution in #1
Full Changelog: v0.19.1...v0.20.0
mcpscan-cli v0.19.1
Official PyPI release for mcpscan-cli (the mcpscan static security scanner), published as mcpscan-cli after PyPI's project-similarity check blocked the name mcpscan (too close to the existing mcp-scan project). The CLI binary is still invoked as mcpscan.
mcpscan v0.19.0 - First PyPI Release
First release of mcpscan published to PyPI: pip install mcpscan.
What mcpscan does
A static security scanner for MCP servers and Claude Code projects, built entirely on the Python standard library.
- Zero runtime dependencies. No third-party packages to audit, no supply chain to inherit, runs in locked-down CI without an install step beyond the scanner itself.
- Static, pattern-based detection across command injection, tool poisoning (prompt injection hidden in tool metadata), hardcoded secrets, over-broad permissions/scope, unsafe deserialization, SSRF, insecure transport, and more — 19 rules total (
mcpscan --list-rules). - SARIF 2.1.0 export, uploads directly to GitHub Code Scanning alongside first-party scanners.
--discover: scans every known local MCP client config on a machine (Claude Desktop, Claude Code, Cursor, VS Code, Windsurf) in one pass, closing the 'which MCP servers am I actually running' blind spot.
This release
- Publishes to PyPI for the first time via Trusted Publishing (OIDC, no stored token), triggered by this release through
.github/workflows/pypi-publish.yml. - README Quickstart now leads with
pip install mcpscan. - Consolidates release automation onto a single publish workflow (removed a duplicate tag-push path).
See CHANGELOG.md for full version history.
v0.18.0 - MCP022: loopback bridges with no Origin check
What's new
MCP022 — a server bound to 127.0.0.1/localhost only still accepts cross-origin requests because it never checks the Origin header.
Binding to loopback is often (wrongly) treated as equivalent to "unreachable by web content." Any browser tab open on the same machine can still reach it via CORS or an unauthenticated WebSocket upgrade. This pattern has now surfaced three times independently: Ollama's long-known default-CORS behavior, BraveMCP's own HTTP bridge before it was fixed in v0.2.0, and Cline's currently-open Hub dashboard WebSocket bug.
Fires on a loopback-bound server (not bind-all — that's MCP018) alongside either a permissive CORS setup (bare cors(), bare CORS(app), or a literal wildcard header) or a WebSocket connection handler with no Origin check.
See the README and CHANGELOG for details.
151 tests passing (was 141). Dogfood self-scan clean.
v0.17.0
Added
MCP021: a click handler that dispatches a privileged action with no event.isTrusted check.
Claude for Chrome's own click-spoofing flaw is the confirmed root cause: a content-script click handler reads data-task-id off the clicked element and dispatches the associated agentic task, but never checks event.isTrusted. Reported to Anthropic's bug bounty program 2026-05-21, confirmed still broken against the shipped v1.0.80 on 2026-07-07 by Manifold Security (corroborated by Bleeping Computer and The Hacker News), no CVE assigned, unfixed as of the most recent reporting found. Any other script with DOM access on the page — commonly another installed browser extension — can forge the click a privileged action relies on.
New mcpscan/rules/dom_trust.py: fires on a click/onclick handler that reads element data (.dataset.x/getAttribute(...)) and forwards it to a privileged sink (extension messaging, or an approve/grant/authorize-named call), with no isTrusted check anywhere in the handler. Scoped to JS/TS source files only.
10 new tests, 141 passing (was 131). Dogfood self-scan clean.
See CHANGELOG.md for full details.
v0.16.0
--discover widens to Amazon Q and Cline global configs
CVE-2026-12957 (Amazon Q, Wiz Research, CVSS 8.5, disclosed 2026-06-26) is a
workspace-committed .amazonq/mcp.json auto-executing with zero trust
prompt. That exact path is already caught by a normal mcpscan directory
scan today (basename matching is directory-agnostic), so this release closes
the real remaining gap: --discover's known-location table never checked
Amazon Q's separate global config (~/.aws/amazonq/mcp.json) or Cline's
global config (VS Code globalStorage) — both live outside any project tree
and were invisible to every mcpscan mode until now.
--discover now checks 7 known locations (was 5). 131 tests passing (was 130).
See CHANGELOG.md for details.
v0.15.1
Fixed
- MCP002's
exfiltratalternative was a bare substring match with no
directive context, unlike every other phrase in the injection regex
(all of which require an actual command shape: "ignore previous
instructions", "do not tell the user", "before X read Y"). This meant
any comment or docstring merely discussing exfiltration, a normal
security term, not an injected instruction, false-positived at MEDIUM
severity. Found while dogfooding mcpscan against secops-toolkit-mcp,
whoseassess_shell_commandmodule comment ("well-documented
destructive/exfiltration shapes") tripped it. Narrowed to two directive
shapes that still catch real injected instructions: an imperative adverb
before the verb (silently/quietly/secretly/covertly exfiltrate), or the
verb followed by an object and a destination (exfiltrate the contents to <host>). 3 new regression tests confirm the false positive is gone
and both real-attack shapes still fire. 130 tests passing (was 127).
Full diff: v0.15.0...v0.15.1