Secure every node in your LLM pipeline
Multi-step pipelines expand the attack surface. Repellox sits at every context handoff, checking what moves between nodes before the next model call.
Three pipeline archetypes, one protection layer
Pipeline structure changes the attack surface. Repellox adjusts inspection to each archetype.
RAG Pipelines
Retrieval-augmented generation pipelines place external documents in model context at query time. Every retrieved chunk is an injection surface. Repellox checks each chunk before it enters the context window, blocking adversarial content in indexed data.
- Inspect chunks before context injection
- Set a confidence threshold for each retrieval source
- Metadata-based source trust tiers
Multi-Step Reasoning Chains
Chain-of-thought and reasoning pipelines move intermediate outputs between steps. Contaminated context can quietly alter later steps. Repellox intercepts each chain boundary, keeping inspection active through the reasoning sequence.
- Per-step context inspection
- Track context propagation across steps
- Score anomalies across the chain
Fully Autonomous Agents
Autonomous agents with long-running loops and persistent memory pose the highest pipeline risk. Context poisoning early in a turn can influence every later action. Repellox keeps a security context thread through the agent lifecycle.
- Lifecycle-wide context integrity tracking
- Memory write authorization rules
- Cross-turn anomaly aggregation
How the protection model works
Three coordinated layers preserve security integrity throughout pipeline execution.
Per-Node Inspection
Each pipeline node registers an inspection checkpoint with the Repellox client. Context entering and leaving nodes is checked independently, so a compromise at one node cannot silently reach the next.
Context Propagation Tracking
Repellox gives context a fingerprint at pipeline entry. As it moves through nodes, the fingerprint records mutations and builds risk scores. Suspicious accumulation prompts escalated inspection at the next node boundary.
Aggregate Audit Trail
Every node inspection event enters a correlated audit log. Replay any pipeline execution, review inspection at each node, and see what caused a block. Export in JSON or send directly to your SIEM.
Inspect your first pipeline in under 10 minutes
Add inspection checkpoints with two API calls. The free tier requires no credit card.