Chagu: Graph-Powered LLM Shielding and Anomaly Detection for Secure AI Pipelines
summary
Chagu claims a graph-native defense that traces prompt→response execution paths, detects prompt injections and semantic drift, and raises alerts in real time by storing events and embeddings in Neo4j Aura and querying with Cypher queries. The demo builds a small graph of users, prompts, models and toolcalls, computes semantic distance and novelty with embeddings plus a one‑class classifier and rule set, uses GDS (shortest paths, centrality) for anomaly fusion, and reports precision/recall/F1 and accuracy on the sample dataset.
AI-written summary · about
chapters (15)
- 0:00Introduction to Chagu
- 1:30What is Chagu and its purpose
- 2:55Problems with agentic AI systems
- 3:38Graph-based defense approach
- 5:02Chagu architecture overview
- 6:26Detection mechanisms in Chagu
- 7:26Visualization and alerting
- 8:07Use cases and examples
- 9:37Workshop introduction
- 10:22Setting up Neo4j in demo
- 11:49Demo walkthrough
- 13:38Executing anomaly detection
- 16:09Q&A and feedback
- 17:15Future of graph-based security
- 21:53Closing remarks and contact
related talks
48mHunting Stealth Adversaries with Graphs & AI
53mLeveraging graph AI and GPUs to win the US Cyber Command AI challenge
20mAdvanced GraphQL Security with AI-Driven Threat Detection
16mGraph-Powered Compliance: Building Intelligent Regulatory Mapping with AI and Neo4j
35mMulti-Layered Guardrails for Cloud Native AI: Enforcin...
30mCybersecurity Risk Assessment Using LLM Agents and Graph Data Science
37mAGENTS OF S.E.A.L.E.D: AI Agentic Cybersecurity Framework
21mEnhancing Threat Intelligence Using AI