Developer Builds 'SentinelCell' Middleware to Shield LangGraph AI Agents from Prompt Injections

A software developer has built SentinelCell, an open-source middleware system designed to protect autonomous LLM agent pipelines from security threats including prompt injection, semantic drift, and cascading hallucinations. The tool operates as a transparent sidecar proxy that intercepts, analyzes, and repairs multi-agent communications in real time without disrupting the pipeline. Unlike traditional firewalls that check syntax, SentinelCell uses a LangGraph-powered state machine to evaluate intent, applying both deterministic string-cleaning and LLM-based semantic repair when payloads are found to be corrupted or malicious. Suspicious or dangerous packets are not simply blocked but quarantined in a Redis-backed Dead Letter Queue, preserving logs and context for later forensic review. The system adopts a fail-closed Zero Trust model, meaning all incoming payloads are treated as untrusted by default and dropped immediately if a deobfuscated injection attempt is detected.