Skills computer-use-agents
📦

computer-use-agents

Content revision r1 High Risk ⚡ Contains scripts⚙️ External commands📁 Filesystem access

Build Secure Computer Use Agents

Computer use agents need safe designs before they control screens, keyboards, and files. This skill provides implementation patterns, sandbox guidance, and Claude tool examples for safer GUI automation.

Supports: Claude Codex Code(CC)
⚠️ 38 Poor

Install with my Agent

Copy this request to your Agent. It includes the canonical Skill page and manifest.

Agent request
Review the Skillstore skill "computer-use-agents" from https://skillstore.io/skills/sickn33-computer-use-agents.md and its manifest at https://skillstore.io/api/skills/sickn33-computer-use-agents/manifest. Verify the artifact. Stop and obtain explicit user consent before installing or changing files.

Your Agent should still show its plan and request any confirmation required by the security policy.

Agent-readable resources

Use these links when an AI agent, crawler, or script needs clean context instead of reading the full page.

Markdown detail GET /skills/sickn33-computer-use-agents.md Signed manifest GET /api/skills/sickn33-computer-use-agents/manifest Signed lockfile GET /api/skills/sickn33-computer-use-agents/lockfile

Test it

Using "computer-use-agents". I need a safe architecture for a desktop automation agent.

Expected outcome:

  • A step-by-step agent loop with observation, planning, action, and feedback stages.
  • A sandbox checklist covering network limits, scoped files, credential isolation, and resource caps.
  • A list of approval points for screenshots, shell commands, and sensitive forms.

Using "computer-use-agents". Review my planned computer-use workflow before implementation.

Expected outcome:

  • A risk summary for screen capture, GUI actions, shell access, and file editing.
  • Specific safeguards such as non-root execution, private temporary files, and allowlisted endpoints.
  • Recommended tests for runaway loops, failed clicks, privacy leaks, and unexpected model actions.

Security Audit

High Risk
v5 • 7/9/2026 Open versioned report

The skill is documentation and sample code for computer-use agents, not an executable package. I confirmed the screen-capture and temporary screenshot handling risks because the examples capture display contents and store them in /tmp. Markdown code-fence detections, sandbox setup commands, and the role-token finding are false positives; no prompt injection or covert exfiltration intent was found.

1
Files scanned
316
Lines analyzed
3
Review items
0
False positives ignored

Confirmed security concerns (2)

High
SKILL.md:18
Screen capture upload
1. PERCEPTION: Screenshot captures current screen state
The skill presents screenshot capture as the core perception step for agents. This is a real sensitive capability because screenshots may include credentials or private data, even though the document recommends sandboxing.
High
SKILL.md:219
Screen capture upload
- screenshot: Capture current screen state
The skill lists screenshot capture as a computer-use tool capability. This is legitimate for the domain but still creates high privacy risk if used outside an isolated environment.
Capability review items (3)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

High
SKILL.md:299
Python subprocess.run
subprocess.run(["scrot", "/tmp/screenshot.png"])
The example invokes the external scrot binary to capture the display and write an image to /tmp. The command is hardcoded, but it is still a sensitive OS-level screen capture operation that requires sandboxing.
Medium
SKILL.md:299
Temp directory access
subprocess.run(["scrot", "/tmp/screenshot.png"])
The example stores a screenshot at /tmp/screenshot.png after capturing the display. Temporary screenshot files can leak sensitive screen contents if permissions or cleanup are weak.
Medium
SKILL.md:301
Temp directory access
with open("/tmp/screenshot.png", "rb") as f:
The example reads the captured screenshot back from /tmp/screenshot.png. This is expected for the workflow, but it handles sensitive screen data in a shared temporary path.

Risk Factors

⚡ Contains scripts (1)
SKILL.md:237-241
⚙️ External commands (7)
SKILL.md:299 SKILL.md:29-85 SKILL.md:85-106 SKILL.md:106-210 SKILL.md:210-235 SKILL.md:235-303 SKILL.md:124
📁 Filesystem access (5)
SKILL.md:125 SKILL.md:128 SKILL.md:129 SKILL.md:299 SKILL.md:301

Detected Patterns

Screen capture upload×2
Audited by: codex View Audit History →
Share & cite this report

Share the versioned assessment report, neutral badge, embed card, and citations. Skillstore reports evidence without deciding whether this Skill is safe.

Open versioned report
Security Assessment

Copy report link

https://skillstore.io/skills/sickn33-computer-use-agents/audits/5?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Markdown badge

[![Skillstore security assessment](https://skillstore.io/badges/skills/sickn33-computer-use-agents/security.svg)](https://skillstore.io/skills/sickn33-computer-use-agents?utm_source=security_passport_badge)

HTML badge

<a href="https://skillstore.io/skills/sickn33-computer-use-agents?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/sickn33-computer-use-agents/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

Embed card

<iframe src="https://skillstore.io/embed/skills/sickn33-computer-use-agents.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>
Academic citations (APA · BibTeX · CFF)

APA citation

sickn33. (2026). computer-use-agents security audit report (audit version 5) [Author version unspecified]. Skillstore. https://skillstore.io/skills/sickn33-computer-use-agents/audits/5

BibTeX citation

@techreport{sickn33-sickn33-computer-use-agents-2026, author = {sickn33}, title = {computer-use-agents security audit report (audit version 5)}, institution = {Skillstore}, year = {2026}, number = {5}, url = {https://skillstore.io/skills/sickn33-computer-use-agents/audits/5}, note = {Author version unspecified} }

CITATION.cff

cff-version: 1.2.0 message: "If you use this Skill, cite its author and this versioned security audit report." title: "computer-use-agents security audit report (audit version 5)" version: "unspecified" type: report authors: - name: "sickn33" date-released: "2026-07-09" url: "https://skillstore.io/skills/sickn33-computer-use-agents/audits/5" identifiers: - type: other value: "skillstore:sickn33-computer-use-agents:audit:5" description: "Skillstore immutable audit report identifier"

Skillstore Score

Why this score Evidence Confidence: Medium
55
Architecture
85
Maintainability
87
Content
69
Community
91
Spec Compliance

What You Can Build

Prototype Desktop Agents

Design the perception, reasoning, and action loop for a new GUI automation agent.

Plan Safe Agent Sandboxes

Map container, network, filesystem, credential, and resource controls before testing automation.

Evaluate Computer Use Tradeoffs

Compare screen-based control risks with browser automation, APIs, and human review workflows.

Try These Prompts

Explain The Agent Loop 
Explain the perception, reasoning, action, and feedback loop for a computer-use agent. Include the main safety risks.
Design A Sandbox 
Design a sandbox for a computer-use agent that needs browser access. Cover network, filesystem, credentials, and resource limits.
Review An Agent Workflow 
Review this computer-use workflow for unsafe screen capture, shell access, file access, and missing human approval points.
Harden Production Controls 
Create a production hardening checklist for a computer-use agent with screenshot tools, shell tools, and text editor tools.

Best Practices

  • Run every computer-use agent inside an isolated desktop with no host credential access.
  • Keep screenshots, shell commands, and file edits behind explicit policy checks and logging.
  • Limit each run by time, steps, network destinations, filesystem scope, and human approval rules.

Avoid

  • Do not give a screen-control agent direct access to a personal workstation.
  • Do not store screenshots in shared temporary paths without cleanup and permission controls.
  • Do not let model-generated actions run shell commands without validation and allowlists.

Frequently Asked Questions

What does this skill help me build?
It helps design agents that observe screens and perform controlled GUI actions through mouse, keyboard, shell, and editor tools.
Does it include production-ready code?
No. It provides patterns and partial examples that need review, completion, testing, and security hardening.
Why is sandboxing required?
A computer-use agent can see private screen data and operate local tools, so host isolation limits damage.
Can I use it with Claude Code or Codex?
Yes. The guidance is useful when planning computer-use workflows with Claude, Claude Code, Codex, or similar tools.
What are the main security concerns?
The main concerns are screenshot privacy, shell access, file editing, credential exposure, and uncontrolled action loops.
Should agents receive unrestricted internet access?
No. Network access should be limited to required endpoints and monitored during every run.

Developer Details

Author

sickn33

License

MIT

Skillstore revision

r1

Version notice

The author did not declare a version.

Repository

https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/computer-use-agents

Ref

3e4b6c31a74a3bd1a291c98cf585d720cb9fbc88

Maintenance freshness

7/18/2026

Usage

36 downloads · 102 views

File structure

📄 SKILL.md