How mxyzptk67/prompt-injection-llm-project works
This repository is a prompt-injection red-teaming and defense application built around a tool-using AI agent. A user request can be classified, routed through either the baseline or defended agent, and returned or blocked; the agent can invoke web search, calculation, and file-reading tools. Separate workflows run curated and dataset-derived attacks, while a Streamlit dashboard displays results and traces. The excerpts support the classifier, guardrail, agent, dashboard, and benchmark relationships shown; some benchmark and reporting modules are indexed but unsampled, so their internal behavior is left unspecified.
Main entry points: defended_agent.py Defended Agent · dashboard.py Security Dashboard · agent.py Baseline Agent · dataset_results_defended.py Dataset Attack Runner
Components 20
Defense
- Regex Classifier
input_classifier.py - Hybrid Classifier
llm_classifier.py - Tool and Output Guards
allowlist.py
Agent Runtime
- Defended Agent
defended_agent.py - Baseline Agent
agent.py - Agent Tools
agent.py
Red-Team Evaluation
- Curated Attack Suite
attack_suite.py - Dataset Attack Runner
dataset_results_defended.py - Benchmark Runner
benchmark.py - Attack Prompts
malignant.csv - Attack Results
attack_results.csv - Benchmark Results
benchmark_results.csv - Defended Dataset Results
dataset_results_defended.csv - Report
report.txt - Notes
notes.txt
Dashboard and Tracing
- Security Dashboard
dashboard.py - Agent Trace
agent_trace.jsonl
Other components
- User
- Ollama Model
- SerpAPI Search
Connections 25
- User → Security Dashboard: submits prompt
dashboard.py - Security Dashboard → Hybrid Classifier: classifies input
dashboard.py - Security Dashboard → Baseline Agent: runs baseline
dashboard.py - Security Dashboard → Defended Agent: runs defended
dashboard.py - Security Dashboard → Curated Attack Suite: loads attacks
dashboard.py - Hybrid Classifier → Regex Classifier: checks first
llm_classifier.py - Defended Agent → Hybrid Classifier: classifies request
defended_agent.py - Defended Agent → Tool and Output Guards: validates tools and output
defended_agent.py - Defended Agent → Baseline Agent: uses agent definitions
defended_agent.py - Baseline Agent → Ollama Model: requests completion
agent.py - Defended Agent → Ollama Model: requests completion
defended_agent.py - Baseline Agent → Agent Tools: dispatches tool calls
agent.py - Defended Agent → Agent Tools: runs allowed tools
defended_agent.py - Agent Tools → SerpAPI Search: searches web
agent.py - Baseline Agent → Agent Trace: writes trace
agent.py - Defended Agent → Agent Trace: writes trace
defended_agent.py - Security Dashboard → Agent Trace: reads trace
dashboard.py - Curated Attack Suite → Baseline Agent: runs attacks
attack_suite.py - Curated Attack Suite → Attack Results: writes results
attack_suite.py - Dataset Attack Runner → Attack Prompts: loads prompts
dataset_results_defended.py - Dataset Attack Runner → Defended Agent: runs prompts
dataset_results_defended.py - Dataset Attack Runner → Defended Dataset Results: writes results
dataset_results_defended.py - Benchmark Runner → Benchmark Results
- Agent Tools → Report: can read file
agent.py - Agent Tools → Notes: can read file
agent.py