What question did this study set out to answer?

The aim is to enhance harmful intent detection in conversations, overcoming limitations of current moderation systems.

March 24, 20260 citationsOpen Access

MURPHY: Conversational Intent Surveillance with Arc-Based Trajectory Detection for Safe Multi-Agent AI Systems

GSGunjan SarkarNational Institute of Oceanography

Key Points

The aim is to enhance harmful intent detection in conversations, overcoming limitations of current moderation systems.
Developed MURPHY, a conversational safety intelligence layer.
Utilized arc-based trajectory detection to evaluate complete conversation arcs.
Implemented a self-amortising cost model for efficient processing of conversation patterns.
Created an oracle-governed pattern learning system for validating new safety patterns.
Achieved 90% detection of threats at nearly zero cost in production environments.
Enabled efficient monitoring of multi-agent AI interactions with reduced user disruptions.

Abstract

Static content moderation systems — regex blocklists and per-message classifiers — fail to detect harmful intent that emerges gradually across a conversation. A user who opens with benign questions and progressively steers toward harmful territory bypasses per-message filters trivially. We present MURPHY (Multi-layer Understanding of Risk through Phased Harm Yielding), a conversational safety intelligence layer that addresses this gap through arc-based trajectory detection. MURPHY evaluates the full conversation arc rather than individual messages, computing a deterministic arc score from four sub-dimensions: topic drift, escalation velocity, sensitive term density, and pattern similarity. MURPHY makes three primary contributions: 1. Arc-based trajectory detection: a deterministic scoring formula that evaluates the full conversation arc across four dimensions, catching slow-burn attacks invisible to per-message classifiers. 2. Self-amortising cost model: a two-tier pattern memory architecture where early LLM inference builds a reusable pattern library, progressively replacing expensive calls with near-zero-cost DynamoDB lookups. At production scale, 90% of threats are caught at near-zero cost. 3. Oracle-governed pattern learning: an external validation loop that approves new safety patterns against historical false-positive rates before activation, preventing paranoid degradation. MURPHY is designed as a surveillance layer for multi-agent AI production systems, operating asynchronously with a command authority agent to avoid blocking user-facing requests. We demonstrate MURPHY through its integration in RRQ, an autonomous YouTube production system, and show how the architecture generalises to any multi-agent deployment requiring safe human interaction. Related publications:- SAGE: Self-Adapting Governance Engine — https://doi.org/10.5281/zenodo.19116770- ILP: Instruction Layer Protocol — https://doi.org/10.5281/zenodo.19101583- RRQ Open Architecture — https://github.com/GunjanGrunge/rrq

Perguntar à IA

Bookmark

View Full Paper