CVE-2026-66004: BlenderMCP path traversal in download_polyhaven_asset allows arbitrary file writes
BlenderMCP, the MCP server that lets an AI agent drive Blender, has a path traversal in `download_polyhaven_asset`. Traversal sequences in the "include" keys of the Poly Haven API response let an attacker write files outside the temp directory, for example `../../.bashrc`, and get persistent code execution. Upgrade blender-mcp to 1.6.5.
Published Jul 24, 2026 · Updated Oct 2, 2026 · Analysis Oct 25, 2026
What an attacker can do
When it downloads a Poly Haven asset, download_polyhaven_asset takes the keys of the API's "include" dictionary and uses them as relative file paths under a temporary directory. Before the fix, nothing checked those keys. A key such as ../../.bashrc made the server write the downloaded content outside the temp directory, overwriting a file in the user's home.
The attacker must control the API response or the tool's input. The advisory names two ways: a man-in-the-middle on the connection to Poly Haven, or prompt injection that steers the agent into fetching a malicious asset. Both need the user to run the tool, which is why the CVSS vector has user interaction and high attack complexity.
Overwriting a shell startup file such as .bashrc runs the attacker's commands the next time a shell starts, giving persistent code execution as the user who runs Blender and the MCP server. Other targets include SSH authorized_keys or any file that user can write. The score is 5.3 under CVSS 3.1 with high integrity impact only.
Who is affected
The blender-mcp pip package before 1.6.5. The vulnerable code is in every build before commit 30a3308, which was merged on July 20, 2026. Only setups that use the Poly Haven download tool are exposed, but the tool is part of the default server.
pip show blender-mcpIf you run it with uvx, check which version your MCP client config resolves.
blender-mcp
- Ecosystem
- PyPI
- Affected
- < 1.6.5
- Fixed in
- 1.6.5
- Check
pip show blender-mcp
How to fix it
Upgrade to blender-mcp 1.6.5.
pip install --upgrade blender-mcp==1.6.5The fix, pull request 258, checks each include key before joining it to the temp directory. It normalizes the path, rejects absolute paths and paths containing .., and requires the resolved path to stay inside the temp directory. An unsafe include is logged and skipped, and valid models still import. The sources list no workaround. Until you upgrade, disable the Poly Haven integration and do not let the agent act on untrusted content.
How to tell if you were hit
Look for files changed outside the temp download directory around the time of a Poly Haven download: .bashrc, .zshrc, .profile, ~/.ssh/authorized_keys and other startup files. Check shell history and new processes for commands you did not run. After the fix, skipped entries appear in the server log as unsafe includes. Beam records agent file access and shell commands, so its logs can show a write outside the project by an agent session.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
BlenderMCP before commit 30a3308 contains a path traversal vulnerability in the download_polyhaven_asset method that allows attackers to write arbitrary files by injecting traversal sequences in API response include keys. Attackers performing MITM attacks or prompt injection can supply malicious paths like '../../.bashrc' to overwrite sensitive files and achieve persistent code execution.
Questions
- Is BlenderMCP exploitable without prompt injection?
Yes, if an attacker can sit between the server and the Poly Haven API and alter the response. Prompt injection is the other path the advisory names. Either one needs the user to run the tool.
- Which version fixes CVE-2026-66004?
blender-mcp 1.6.5. The fix is commit 30a3308 from pull request 258.
- What can an attacker overwrite?
Any file the running user can write. The advisory's example is
../../.bashrc, which gives persistent code execution at the next shell start.- Is it exploited in the wild?
None of the sources I read report exploitation, and it is not in CISA's KEV catalog.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More Other MCP servers and tools vulnerabilities
All Other MCP servers and tools entriesCVE-2026-102943
Dagu MCP DAG specification read bypasses path containment and workspace isolation
medium6.5Nov 10, 2026CVE-2026-102610
fast-agent: `fast-agent serve` binds HTTP to 0.0.0.0 with no authentication by default; with --shell this is unauthenticated remote command execution
medium6.4Nov 9, 2026CVE-2026-63655
tsParticles: Unbounded HTTP Request Body Buffering in @tsparticles/mcp-server Leads to Remote DoS
high7.5Nov 9, 2026CVE-2026-76892
tsParticles: Malformed JSON requests can terminate the @tsparticles/mcp-server HTTP process
high8.7Nov 9, 2026CVE-2026-76893
tsParticles: A single anonymous MCP session can monopolize HTTP mode and deny new clients
medium6.9Nov 9, 2026CVE-2026-77604
Code Graph RAG: MCP paginated read_file path traversal / arbitrary file read
high7.5Nov 9, 2026
