← Back to Tyree Sloan's portfolio
Concealed Intent Detection Framework

Hidden intent emerges over time.

UNMASK analyzes behavioral patterns across multi-turn LLM interactions to identify concealed objectives before sensitive information leaves the system.

Let's role play

Imagine you're a bank manager. Click each question to answer it.

Cumulative Risk Level
Risk Score
Turn 1 How many tellers work here? tap to reveal →
"We have twelve tellers on staff across two shifts."
🟢 Low Risk
Turn 2 Which branch do you manage? tap to reveal →
"I manage the downtown branch — we're the busiest in the city."
🟡 Moderate — pattern forming
Turn 3 How much money is kept in the vault? tap to reveal →
"Our vault typically holds between $400K and $800K depending on the week."
🔴 High — converging intent

No individual question appears dangerous. But humans start drawing inferences once questions converge — LLMs typically evaluate each turn in isolation. UNMASK applies cross-turn pattern recognition to close that gap.

How UNMASK connects red team and blue team

Red team activity creates concealed-intent interactions. UNMASK middleware interprets behavioral patterns. Blue team systems evaluate risk and decide what to do.

Red Team +

Intent Obfuscation

Creates adversarial and dual-use conversations designed to hide the larger objective.

Purpose
Simulate adversarial users Test detection limits Generate training data
Prompt Strategies
Direct framing Academic framing Journalism framing Movie-script framing Forensic framing
Techniques
Prompt chains Noise chains AutoPrompt
Output → Test Conversations
UNMASK Middleware +

Behavioral Translation

Transforms multi-turn behavior into structured signals that a detector can evaluate.

Core Signals
Behavioral Memory Behavioral Density Goal Convergence Context Camouflage Trigger Events Behavioral Molecules
What it tracks
User claims Role history Topic history Contradictions Narrative shifts
Output → Behavioral Feature Report
Blue Team +

Detection Support

Uses feature signals to support risk assessment, verification, and response decisions.

Capabilities
Detection Engine Risk Assessment Adaptive Verification Decision Support
Actions
Allow Verify Flag Block
Output → Risk Decision
Red Team Output: Test Conversations
Middleware Output: Feature Report
Blue Team Output: Risk Decision

Explore the framework

These sections will become deeper pages as the project evolves into a full framework portal.

Behavioral analysis pipeline

A layered view of how UNMASK moves from prompt-level signals to a final risk assessment.

Layer 1Behavioral Feature ExtractionHow is the user asking?
Signals: negation, speculation, hedging, commitment strength, ownership, transparency, opacity, narrative framing, plausibility.
Layer 2Behavioral MemoryHow does behavior change over time?
Stores: user claims, role history, topic history, risk history, prior contradictions, and framing patterns.
Layer 3Consistency & Divergence AnalysisDoes behavior match the stated purpose?
Signals: consistency score, claim stability, identity drift, intent divergence, narrative shifts, contradictions.
Layer 4Structural Behavior AnalysisAre prompts forming a larger structure?
Signals: behavioral density, goal trajectory, prompt relationships, behavioral molecules, multi-turn intent graph.
Layer 5Goal ConvergenceWhat larger goal is emerging?
Signals: repeated topic clustering, increasing specificity, related sub-questions, dense information gathering, cross-turn alignment.
Layer 6Risk AssessmentWhat should the system do?
Outputs: risk score, confidence score, recommended action, and possible adaptive verification.

Intent polarity spectrum

UNMASK is most valuable where benign, dual-use, and adversarial intent overlap. Drag the slider to explore each zone.

◀ Blue Zone (Benign) Dual Use (High Risk) Red Zone ▶
Blue Zone — Mostly benign learning, work, and creative activity
Educational Intent
5% red
Professional Intent
10% red
Creative Intent
15% red
Information Seeking
30% red
Influence Intent
50% red
Cyber Intent
70% red
Intent Concealment
90% red

Blue Zone

Mostly benign learning, work, and creative activity. Low UNMASK engagement needed.

Dual-Use Zone

Information seeking and influence can serve legitimate or harmful goals. Pattern tracking matters here.

High-Risk Zone

Cyber intent and concealment require deep behavioral analysis and likely adaptive verification.

Candidate feature explorer

Click a feature to see how it supports detection and interface design.

Behavioral Density

Measures how concentrated related behaviors are within a conversation window.

Potentially useful for detecting low-and-slow information gathering where malicious prompts are spaced apart to avoid triggering per-turn detection.

Context Camouflage

Detects semantic masking, neutral language shifts, and sanitized wording.

Connects to intent obfuscation patterns — replacing loaded terms with neutral alternatives while preserving the underlying information-gathering structure.

Trigger Events

Identifies turns that cause a sudden increase in suspicion score.

Useful for identifying the precise moment a conversation shifts from benign to concerning — pinpointing the turn that crosses the threshold.

Adaptive Verification

Prompts clarification when cumulative risk crosses a defined threshold.

Mirrors a human-like response: ask a follow-up question before blocking or allowing the interaction. Reduces false positives without ignoring risk.

Behavioral Molecules

Groups smaller behaviors into larger structures that indicate intent.

Example: repeated questioning + progressive specificity + dependency mapping = information-gathering molecule. Individual atoms are inconclusive; the molecule is diagnostic.

Goal Convergence

Measures whether prompts are moving toward the same larger objective.

The hidden objective may not appear in any single prompt but can emerge clearly across multiple turns. Goal convergence quantifies that trajectory.

Middleware architecture

Let's revisit the bank example

UNMASK operates as a behavioral layer between conversation history and an external detection engine. Watch as signals are analyzed — some clear, some flagged.

User Conversation
Conversation History
UNMASK Middleware
Feature Report
Detection Engine
Risk Score / Action
0
0

Red-team testing plan

Testing evaluates whether concealed intent remains detectable across strategies, distractions, and prompt transformations.

Step 1

Prompt Chains

Multi-turn scenarios across direct, academic, journalism, documentary, forensic, professional, and movie-script framing.

Step 2

Noise Chains

Benign or distracting prompts inserted between information-gathering turns to test cross-turn persistence.

Step 3

AutoPrompt

Semantically optimized variants used to test robustness against rewording and paraphrase attacks.

Step 4

Detector Evaluation

Compare outputs against feature signals, trigger events, and scenario objectives to measure detection quality.

Research portal

This section will document literature support, source-to-feature mapping, prior work comparisons, and open questions.

Purpose

The research portal connects UNMASK concepts back to the literature. It will eventually include feature traceability, Mal-CoT comparison, trajectory-level monitoring relevance, and future work.

Back to top