Back to skill

Security audit

Meshy 3D Agent

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious with medium confidence because it reports an unapproved global CLI installation, an API-key option that conflicts with the stated sign-in method, and missing disclosures before sending user content to Meshy.

Consider avoiding installation until the global install behavior is changed to require approval, the API-key authentication option is clearly disclosed or removed, and users are told before descriptions, images, or model files are sent to Meshy. The verdict is suspicious, not a claim that the author acted maliciously. Scanning was partial: all eight listed components were scanned, but several local references could not be resolved, so the evidence is incomplete.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (14)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
91% confidence
Finding

Checking the account balance can reasonably support Meshy operations and is not itself a mismatch. However, the explicit --api-key-file ./.env option introduces local credential-file access and API-key authentication inconsistent with the stated sign-in workflow. The chunk does not establish unrelated data access or exfiltration.

Content

No source excerpt is available for this finding.

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/creative-lab.md (reported line 46)May include surrounding context.

bash
meshy creative-lab figure build create --input-task-id PROTOTYPE_ID --async --format json --no-update-check
meshy creative-lab lamp build create --input-task-id PROTOTYPE_ID --model-format stl --options '{"diameter_mm":180,"light_source_preset":"none"}' --async --format json --no-update-check
meshy creative-lab keychain build create --input-task-id PROTOTYPE_ID --model-format obj --options '{"size_mm":50}' --async --format json --no-update-check
meshy creative-lab fridge-magnet build create --input-task-id PROTOTYPE_ID --model-format obj --options '{"size_mm":60}' --async --format json --no-update-check
meshy creative-lab keycap build create --input-task-id PROTOTYPE_ID --candidate-id CANDIDATE_ID --async --format json --no-update-check
meshy creative-lab fidget-pixel build create --input-task-id PROTOTYPE_ID --async --format json --no-update-check

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/setup.md (reported line 238)May include surrounding context.

so verify that branch with a free balance call instead:

bash
meshy balance --api-key-file ./.env --format json --no-update-check

The CLI reads only the key from that file. Do not source or display it. Never ask for an API

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · references/troubleshooting.md (reported line 34)May include surrounding context.

md
- **Network failure during verification**: say it is a connectivity problem, keep the stored
  credential, and retry the check when the network is back. Do not start a new authorization.
- **Auth rejection**: use the login flow in [setup.md](setup.md). Do not clear all profiles.
- **Refresh failure**: a still-valid access token may keep working; an expired one cannot.
  Check connection/server errors before asking the user to authorize again.
- **Login code expired or denied**: rerun the same `auth login --device` for a fresh code, and
  show the new URL and code. Nothing paid is retried because a login timed out.

The text-to-motion workflow submits the user's motion description to Meshy without a privacy disclosure.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow instructs the agent to run meshy text-to-motion create with a motion description, transmitting that content to a third-party API. The file explains output formats and limits but does not warn that the description leaves the user's environment.

Content

No source excerpt is available for this finding.

The photo-based workflow sends user image URLs to Meshy's API without a privacy warning.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow instructs the agent to submit INPUT_IMAGE to Meshy for prototype generation, potentially exposing personal photos and their URLs to a third-party service. Although API processing is mentioned, the file does not warn users that their images will be shared with an external provider.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/delivery.md (reported line 52)May include surrounding context.

meshy balance --format json --no-update-check

text

Spend rules: work already authorized by the user proceeds without asking again, stage by stage,
to the end of the agreed plan. Ask only for a **new** cost the user has not agreed to — an extra
stage, a second variant, a rerun after a disappointing result. Exit 9 (`insufficient_credit`)
stops new paid submissions immediately: keep everything already produced, report what is

Local model files are submitted to Meshy without a warning that their contents leave the user's system.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow instructs the agent to pass a user-selected local model, or a previously delivered local model, to Meshy's processing commands through --model-url. It does not disclose that using this fallback uploads the asset to a third-party service, which matters for private or proprietary models.

Content

No source excerpt is available for this finding.

Reference-image workflows send user files to Meshy without a privacy disclosure.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow instructs the agent to submit a user-provided photo to Meshy's API, but does not require informing the user that their image will leave the device and be processed by a third-party service. Approval of the generation route and its cost does not disclose this privacy consequence; the same omission affects the multi-image and image-editing recipes.

Content

No source excerpt is available for this finding.

The workflow sends user-provided images to Meshy without a user-facing disclosure that local files will leave the device.

Medium
Category
Output Handling
Confidence
87% confidence
Finding

The recipe instructs the agent to pass a user's photo, including a local file, directly to Meshy's image-to-3d service. It does not instruct the agent to warn that the image will be uploaded to a third-party service, which matters for private or sensitive photos; the sketch-upload route has the same omission.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

The skill explicitly directs the agent to install a global npm package without approval, creating persistent changes outside the job's workspace and executing downloaded package code. Installing a dependency is relevant to this workflow, but a request to generate a model does not necessarily authorize global software installation; prohibiting sudo limits privilege escalation but does not eliminate user-level supply-chain or environment-modification risks.

Content

Scanner excerpt · references/setup.md (reported line 44)May include surrounding context.

npm install -g meshy-cli@0.8

text

Check and install without asking first; say in one line that you are installing it. Install at
most once: when `meshy --version` still reports a version older than 0.8.5 afterwards, the
registry this machine uses lags npm, so report that instead of installing again. Never use `sudo`
or change system policy: when the install is refused (permissions, a managed machine, a sandbox), use the temporary package

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/setup.md (reported line 115)May include surrounding context.

md
with the host's own tool for reading a running command, until it exits. Browser approval ends
the CLI process, not your turn, and a process still running when the turn ends has nobody to
read its result. So do not detach it (`&`, `nohup`, a new terminal), do not end the turn with
"authorize and I will continue", and do not ask the user to reply "done". A tool yield or timeout
is not a code expiry. While it polls, do not start a second login, kill it or poll the API
yourself; send a brief progress note during a long wait. A host whose own skill says it wakes the
session when a background command exits follows that procedure instead.

Overly Broad Trigger

Low
Category
Trigger Abuse
Confidence
75% confidence
Finding

Overly Broad Trigger: description clause 'Use when the user asks to make, texture, rig, animate or convert a 3D model, or wants something 3D-printed' activates on 'make', a common word that will activate in many unintended contexts

Content

No source excerpt is available for this finding.

The sign-in workflow explicitly copies a device code into chat despite the security manifest's promise never to read, print, or copy one.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

L040-L041 state that the skill never reads, prints, or copies a device code. However, L109-L112 instruct the agent to read the code from CLI output and send it to the user verbatim in chat. Although this supports the documented browser sign-in flow, it directly contradicts the stated credential-handling guarantee.

Content

No source excerpt is available for this finding.