Back to skill

Security audit

scientific-drawing-skill-1-0-0

Security checks for vulnerabilities and agentic risk

Overview

The skill is for scientific figure generation, but its installation and editor defaults expose users to unreviewed remote code, broad network access, and under-specified telemetry.

Review this carefully before installing. Use a virtual environment or container, pin the external repository to a reviewed commit, avoid running the editor on 0.0.0.0 unless you have authentication and network controls, and do not submit sensitive unpublished research or proprietary images unless you accept the third-party AI provider and telemetry implications.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (3)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:81
Finding

Unpinned Dependencies and Mutable Remote Source Installation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 81–94; related dependency metadata at line 16 and upgrade instructions at lines 406–407
Vulnerability Type: Software supply-chain exposure through mutable remote sources and unpinned dependencies
Risk Level: High

Vulnerable Snippet:

bash
git clone https://github.com/ResearAI/AutoFigure-Edit.git

cd /home/davidzhao/.openclaw/workspace/skills/scientific-drawing-skill

pip install -r requirements.txt

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install transformers datasets accelerate
pip install opencv-python pillow scikit-image
pip install lxml svgwrite svgpathtools
pip install flask flask-cors

The update instructions also execute an unconstrained dependency upgrade:

bash
cd /home/davidzhao/.openclaw/workspace/skills/scientific-drawing-skill
pip install -r requirements.txt --upgrade

Technical Analysis

The installation process clones the default branch of an external Git repository without selecting a reviewed commit, signed tag, or verified release artifact. Consequently, the downloaded implementation can change after this skill has been audited.

The documented pip commands also omit exact versions and cryptographic hashes. The referenced requirements.txt is not present in the audited project, so its contents and transitive dependencies cannot be reviewed. Python package installation may execute package build backends and installation hooks. A compromised upstream repository, package release, account, or transitive dependency could therefore introduce executable code during installation.

This combines mutable remote payload retrieval with insecure dependency management. HTTPS protects data in transit but does not establish that the current upstream content is the same content that was previously reviewed.

Attack Path

  1. An attacker compromises the referenced reposito ...[truncated 1116 chars]
Remediation
View remediation

Remediation Suggestions

  • Pin the external repository to a reviewed full commit hash or a cryptographically signed release tag.
  • Record and verify the expected commit identity before installation.
  • Include the implementation and dependency manifest in the auditable project rather than relying on undocumented remote state.
  • Pin every direct and transitive dependency to an exact version.
  • Generate a lockfile and require package hashes, such as through pip install --require-hashes.
  • Remove unconstrained --upgrade instructions from normal installation and maintenance procedures.
  • Use an isolated virtual environment or container with minimal filesystem and network permissions.
  • Perform dependency vulnerability, provenance, and package-name checks before publishing updates.
  • Prefer trusted package indexes explicitly and prevent fallback to unintended indexes.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:189
Finding

Upload-Capable SVG Editor Exposed on All Network Interfaces

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 189–198 and 237–243
Vulnerability Type: Insecure network service configuration
Risk Level: Medium

Vulnerable Snippet:

text
--port: server port (default: 8080)
--host: bind host (default: 0.0.0.0)
--data_dir: SVG file storage directory
bash
scientific_drawing_edit_server --port 8080
yaml
editor:
  port: 8080
  allow_upload: true
  max_file_size: 50MB

Technical Analysis

The documented default host is 0.0.0.0, which makes the editor listen on every available network interface rather than only the loopback interface. The configuration simultaneously enables uploads.

The documentation does not specify authentication, authorization, TLS, CSRF protection, rate limiting, filename normalization, isolated storage, or SVG sanitization. SVG files can contain active elements and external references, so accepting them without strict sanitization can create stored cross-site scripting or unwanted external-resource loading risks when files are rendered in a browser.

The server implementation is absent from this project, so specific upload-processing defects cannot be confirmed. Nevertheless, the documented network exposure itself is an insecure default: it makes an upload-capable service remotely reachable unless an external firewall blocks access.

Attack Path

  1. A user starts the editor according to the documented configuration.
  2. The server binds to 0.0.0.0:8080, making it reachable from other systems allowed by the host and network firewall.
  3. A remote client connects to the editor and accesses its upload functionality.
  4. If the absent implementation lacks authentication or adequate file validation, the client can upload crafted or oversized content.
  5. Crafted SVG content may affect users who open it in the editor, while repeated uploads may consume storage or application resources.

Impact Asses

...[truncated 518 chars]

Remediation
View remediation

Remediation Suggestions

  • Change the default binding address to 127.0.0.1 or ::1.
  • Require explicit user confirmation before binding to a non-loopback interface.
  • Add authenticated sessions and authorization checks for every upload, read, edit, and delete operation.
  • Place remotely accessible deployments behind TLS and a hardened reverse proxy.
  • Apply CSRF protection to state-changing browser requests.
  • Enforce server-side upload size limits before buffering request bodies.
  • Generate server-controlled filenames and store uploads outside executable or publicly served directories.
  • Parse and sanitize SVG files using an allowlist that removes scripts, event handlers, foreign objects, unsafe URLs, and external references.
  • Add rate limits, storage quotas, audit logging, and restrictive firewall guidance.
  • Document the implemented security controls and include the server source in the project so they can be audited.

other

Note
Location
SKILL.md:491
Finding

Telemetry Collection Lacks Verifiable Consent and Data-Handling Controls

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 491–497
Vulnerability Type: Underspecified collection of usage and system telemetry
Risk Level: Low

Relevant Snippet, translated to English from the source documentation:

text
The skill anonymously collects the following usage data to improve performance:
- Feature usage frequency
- Average processing time
- Common error types
- System configuration information, including CPU, GPU, and memory

Privacy statement: All data is collected anonymously and does not include
personally identifying information or specific content data.

Technical Analysis

The documentation states that the skill collects usage frequency, processing times, error categories, and hardware configuration. It does not identify the telemetry recipient, network endpoint, transmission security, exact event schema, retention period, legal basis, opt-in mechanism, deletion procedure, or configuration for disabling collection.

The project contains no telemetry implementation that would allow the anonymity claim or transmitted fields to be verified. Hardware characteristics, timestamps, error details, and network metadata can also become identifying when combined, even if explicit personal identifiers are omitted.

This finding concerns inadequate disclosure and control rather than confirmed credential or content exfiltration. The available evidence does not establish that telemetry is currently transmitted because the corresponding implementation is absent.

Attack Path

  1. A user installs and invokes the external implementation described by the skill.
  2. The implementation performs the telemetry collection claimed in the documentation.
  3. Usage events and system-configuration attributes are transmitted to or retained by an unspecified recipient.
  4. The user has no documented method to provide informed consent, inspect the event data, disable collection, or request de ...[truncated 666 chars]
Remediation
View remediation

Remediation Suggestions

  • Disable telemetry by default and require explicit, informed opt-in consent.
  • Document the recipient, endpoint, event schema, transmission protocol, retention period, access policy, and deletion process.
  • Provide a clearly documented environment variable or configuration option that completely disables collection.
  • Minimize collected attributes and avoid stable identifiers, precise timestamps, detailed hardware fingerprints, file paths, prompts, generated content, and raw error messages.
  • Prevent API keys, authorization headers, environment variables, and user content from entering telemetry events.
  • Publish the telemetry implementation with the skill so its behavior can be audited.
  • Add automated tests proving that sensitive fields are removed and that the opt-out control prevents all telemetry network requests.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (4)

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 · SKILL.md (reported line 100)May include surrounding context.

3. 环境配置

bash
# 复制环境配置文件
cp .env.example .env

# 编辑.env文件,添加你的API密钥
# GEMINI_API_KEY=your_gemini_api_key_here

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill instructs users to submit method text, reference images, and API keys for processing by external providers such as Gemini or OpenRouter, but it does not clearly warn users near usage that their inputs may leave the local environment. This can cause unintentional disclosure of unpublished research content, proprietary figures, or sensitive data to third-party AI services.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

SKILL.md 前文主要将该技能描述为本地的科研插图生成、风格迁移与交互式编辑工具,但在 L490-L497 又宣称“技能会匿名收集”使用频率、错误类型和系统配置信息。这属于额外的数据收集行为;然而当前给出的技能文件只有依赖、环境变量和使用说明,没有任何与遥测、上报端点或数据收集配置相匹配的实现线索,因此文档声明与可见行为范围存在不一致。

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The README discloses anonymous usage statistics only near the end of the file, after installation and usage sections. Because the skill collects system configuration and usage-frequency data, users should be warned earlier and more prominently before enabling or using the skill.

Content

No source excerpt is available for this finding.