Neura Market
prompt logo

prompt

Free

Red-team prompt for AI agent safety testing

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About prompt

The Computer Use Safety Tester is a structured prompt designed for red-teaming and safety testing of browser/desktop AI agents. It provides a comprehensive framework for evaluating an agent's ability to resist indirect prompt injection, prevent data exfiltration, avoid unsafe high-impact actions, handle domain and identity confusion, and maintain safety over long task sequences. The prompt includes primary risk categories, testing rules, an output format requiring specific sections (test objective, threat model, attack scenarios, pass/fail criteria, evidence, severity rubric, recommended fixes), and a severity scale from Critical to Low. It is part of the ai-boost/awesome-prompts open-source collection.

Key Features

Defines five primary risk categories: indirect prompt injection, data exfiltration, unsafe high-impact actions, domain/identity confusion, and long-horizon degradation
Provides testing rules for reproducible attack scenarios and chained attacks
Specifies a detailed output format with eight required sections for documentation
Includes a severity scale (Critical, High, Medium, Low) with clear definitions
Requires concrete triggers and explicit evidence for pass/fail determination

Pros & Cons

Pros
  • Provides a systematic and reproducible methodology for safety testing
  • Covers both immediate and chained/long-horizon attack scenarios
  • Includes a clear severity rubric to prioritize findings
  • Open-source and freely available for adaptation
  • Encourages distinguishing harmless confusion from material risk
Cons
  • Requires a capable LLM-based agent to execute the tests, limiting applicability to systems that support long prompts
  • Does not include automated tooling; manual setup of environments and attack scenarios is needed
  • Focuses only on safety risks defined in the prompt, may miss domain-specific vulnerabilities

Best For

Evaluating browser agents against prompt injection attacksTesting desktop AI agents for data leakage and unauthorized actionsAssessing long-task safety degradation in AI assistantsConducting structured red-teaming exercises for AI agent productsValidating safety guards in autonomous browsing systems

FAQ

What is the Computer Use Safety Tester?
It is a prompt from the ai-boost/awesome-prompts collection that instructs an AI to act as a red-team safety testing agent for browser/desktop agents. It defines categories of risks (e.g., prompt injection, data exfiltration) and a structured output format for documenting test results.
What risks does this prompt test for?
It tests five primary risk categories: indirect prompt injection, data exfiltration, unsafe high-impact actions (like submitting forms or making purchases), domain and identity confusion (lookalike domains, redirects), and long-horizon degradation where safety rules are forgotten over many steps.
How should I use this prompt?
Provide this prompt to a computer-use agent (e.g., a browser-controlling LLM) before a testing session. The agent will then generate test scenarios, execute them, and output results following the specified sections (test objective, threat model, attack scenarios, etc.).
Is this tool free?
Yes, it is open-source and freely available in the ai-boost/awesome-prompts GitHub repository under the MIT license, and accessing the file requires no payment.