Red-team text prompts for vulnerabilities using adversarial inputs.
npx skills add https://github.com/JGalego/Bag-of-Tricks --skill strawman
Please help me install this Agent Skill. Skill: strawman Source: https://github.com/JGalego/Bag-of-Tricks/tree/main/strawman Command: npx skills add https://github.com/JGalego/Bag-of-Tricks --skill strawman
This Skill enables you to red-team prompts before shipping them, ensuring they are robust against adversarial inputs and potential security breaches.
Run the strawman skill on your system prompt with the following command: strawman my_system_prompt.txt
High-intent search queries and answers about installing and using this skill.
To red-team a prompt for jailbreaks, provide your system prompt in a text file. The skill simulates adversarial attacks like jailbreaks and instruction injection, identifies scope enforcement failures, and provides feedback on weaknesses with suggested fixes for prompt hardening.
Prompt hardening strengthens system prompts against adversarial inputs. It identifies vulnerabilities like scope derailment and security breaches, then applies suggested fixes to ensure the LLM withstands malicious attacks before deployment.
Yes, testing prompt vulnerabilities requires API access to LLM backends like Anthropic, OpenAI, or Gemini. The red-teaming process uses these APIs to simulate adversarial attacks and evaluate your system prompt's responses to malicious inputs.
To check for instruction injection vulnerabilities, run an adversarial attack simulation against your system prompt. This identifies security breaches and scope enforcement failures, providing detailed feedback on weaknesses and hardening recommendations.
Before shipping LLM prompts, test against adversarial attacks including jailbreaks, instruction injection, and scope derailment. Simulating these attacks identifies security breaches and scope enforcement failures, ensuring your prompt withstands malicious inputs.
Yes, this skill supports Google Gemini models. It requires API access to LLM backends like Anthropic, OpenAI, or Google Gemini to simulate adversarial attacks and provide feedback on system prompt vulnerabilities and suggested fixes.
glebisAccess OpenAI, Anthropic, Google Gemini, and Ollama through a unified command-line interface.
Duyen9904Design and optimize production prompts for LLM performance and reliability.
ejolietDesign prompts, evaluate LLM outputs, and build agentic workflows.
bpnrockstarTest LLM applications for prompt injection, jailbreak, and guardrail bypass vulnerabilities.
