ArthaShield runs AI where the stakes are real. Two controls make that safe for regulated institutions: an LLM-as-Judge layer that checks every output, and an agentic security perimeter that bounds what agents can do.
No AI output reaches a user or a workflow unchecked. Every recommendation, classification and summary ArthaShield produces is scored by an independent judge model against a versioned rubric before it is shown or acted on.
A task model produces an output, grounded in evidence from the knowledge graph.
A separate judge scores it on groundedness, policy compliance, scope and PII leakage.
Pass → released with its score. Borderline or fail → quarantined for a human, never auto-actioned.
Judgments are logged and sampled against human reviewers; rubrics ship only through an eval gate.
The judge is a separate model and prompt from the generator — it never grades its own work.
Scoring is against explicit, versioned criteria, so a verdict can be explained and reproduced.
Low-confidence outputs are held for review, not shipped. Safety beats coverage.
Groundedness ties each claim to source data; every judgment is logged with score and rationale.
ArthaShield's agents read context and propose actions, but they operate inside a controlled perimeter. Autonomy is bounded by least privilege, human approval and a complete audit trail.
Each agent is granted only the scoped tools its task needs. Nothing is ambient or standing.
State-changing actions — file a report, update a case, trigger a workflow — require explicit approval.
Instructions come only from the user or task. Content read from documents, tickets and the web is data, never commands.
Tools run in isolated contexts with rate limits and circuit breakers to contain runaway behaviour.
Agent outputs are validated against strict schemas before they ever touch a system of record.
Credentials are never exposed to the model; redaction at the boundary; every action is logged immutably.
These controls are designed to map onto the security and AI-governance standards our customers are audited against.
This page describes ArthaShield's security architecture and approach. Specific certifications and assurances are shared under NDA on request.
