Sentinel operates in four controlled modes:

 

MONITOR – Detects and logs anomalies (no blocking)

GUARD – Blocks only high-confidence attacks

ENFORCE – Enforces strict filtering on all detections

AUDIT – Offline forensic analysis and compliance review

 

Mode selection is deployment-specific.

Organizations typically start in MONITOR mode and progress to GUARD or ENFORCE based on validated performance.

SENTINEL RAG monitors RAG pipelines for:

 

• Poisoning attacks (semantic manipulation)
• Adversarial document injection
• Authority mimicry and consensus attacks
• Context pollution and subtle misinformation

 

Core functions:

• Real-time anomaly detection
• Cryptographic provenance verification (HMAC-signed documents)
• Authority-weighted retrieval control
• Secure retrieval filtering
• Audit signal generation

 

Evaluation results (controlled conditions):

• Detection rate: 100% (24/24 attack scenarios)
• False positives: 0 (500 legitimate documents)
• Overhead: Minimal (<100ms median)

 

Production performance depends on deployment-specific factors.

See whitepaper for threat model and evaluation methodology.

Sentinel sits between retrieval and generation.

 

It does not modify:

• The LLM
• The prompt
• The vector database

 

It controls:

• Document selection
• Trust weighting
• Semantic consistency

 

This makes Sentinel framework-agnostic and vendor-neutral.

SENTINEL RAG is NOT:

 

• An LLM or language model
• A chatbot or conversational AI
• A prompt-injection filter
• A content moderation tool
• Open source software
• A zero-configuration solution
• A replacement for access control
• A guarantee against insider threats

 

This system addresses a specific attack surface:

poisoning of retrieval corpora in RAG architectures.

 

It is not a universal AI security solution.

 

Critical distinction:

Sentinel does not attempt to "fix" poisoned answers.

It prevents poisoned context from reaching the model.

Private access only.

Enterprise pilots under NDA.

 

Contact: info@sentinelrag.com

 

No public demo available.

No API documentation published.

Implementation details disclosed under separate agreement.