Detect prompt injections and jailbreak attempts in LLM inputs.
npx skills add https://github.com/ricable/mcai --skill prompt-guard-ricable
Please help me install this Agent Skill. Skill: prompt-guard Source: https://github.com/ricable/mcai/tree/main/.agents/skills/prompt-guard Command: npx skills add https://github.com/ricable/mcai --skill prompt-guard-ricable
This Skill helps protect LLM applications from prompt injection, jailbreak attempts, and poisoned third-party content that could override instructions or corrupt outputs.
Use the prompt-guard skill to evaluate a prompt for injection risk and block it when the score exceeds your chosen threshold.
High-intent search queries and answers about installing and using this skill.
Prompt injection detection applies transformer-based classification to score user prompts and block suspicious jailbreak attempts before they reach the model. This filters malicious inputs by evaluating injection risk against configurable thresholds.
RAG protection involves inspecting retrieved documents and API responses for embedded instructions or context poisoning. Third-party data screening evaluates scraped text and knowledge base articles to prevent poisoned content from corrupting outputs.
Batch safety checks process many documents efficiently when preparing retrieval corpora or downstream LLM inputs. This allows multilingual chatbots to apply low-latency filtering and score multiple prompts simultaneously for injection risk.
Configurable thresholds allow multilingual chatbots to define acceptable injection risk scores. You set a specific score limit, and the content screening mechanism blocks any user prompt or retrieved document exceeding that chosen threshold.
Low-latency filtering ensures safe deployment without significant delays. Transformer-based classification evaluates input prompts efficiently, allowing real-time user prompt filtering and third-party data screening to maintain responsive chatbot interactions.
Inspecting API responses prevents third-party content from overriding instructions or corrupting outputs. Jailbreak detection screens external data sources for embedded malicious instructions, ensuring that retrieved or scraped text cannot manipulate your LLM application.
PlaidmustacheClassify images and retrieve text matches using CLIP embeddings.
niclasmartinsson-sysSanitize model outputs using a Model Armor template.
gagan114662Generate sentence, text, and image embeddings using pre-trained transformer models.
FelipeMainesInject contextual rules into prompts via an 8-layer pipeline.
