Most campus tours are forgettable. You walk past buildings, maybe glance at a plaque, and leave without understanding why anything matters. The history exists, but it’s buried behind static text and missing context. We wanted to change that. This is the story of how we built Campus AR Heritage Guide, an offline-first augmented reality system that turns physical spaces into interactive, narrative-rich experiences — and how it became a finalist at the Samsung PRISM Metaverse Hackathon.

The Core Idea

What if history wasn’t something you read, but something you see layered onto the real world? Instead of:

• static boards

• guided tours

• PDFs or brochures
  We built a system where:

• you walk near a monument

• your phone understands where you are

• overlays its story directly in front of you
  No internet required.

The Real Problem We Focused On

Most AR demos fail because they assume:

• stable internet

• controlled environments

• high-end devices
  Real campuses don’t have that luxury.

Constraints we designed for:

• poor or zero connectivity

• outdoor environments with inconsistent lighting

• users who don’t want friction
  So we reframed the problem:

Can AR work reliably offline, in the real world, at scale?

What We Built

Campus AR Heritage Guide is a native Android application that combines:

• geospatial AR

• sensor fusion

• offline intelligence

• voice interaction
  to create a seamless exploration experience.

How the Experience Feels

Step 1: Open the app

You see nearby monuments sorted by distance.

Step 2: Walk around

As you move, the system updates in real time.

Step 3: Enter AR mode

Point your camera and:

The monument comes alive with contextual overlays.

Step 4: Ask questions

You can ask:

• “Who built this?”

• “Why is this important?”
  And get answers — even offline.

Why Offline-First Was Non-Negotiable

Most AR apps break without internet. We engineered the opposite.

The app should work perfectly offline, and improve when online.

System Architecture (High Level)

Core Technical Components

1. Geo-Spatial AR using ARCore

We used Google ARCore’s Geospatial API to:

• anchor content to real-world coordinates

• maintain stability across movement

• ensure overlays stay fixed to physical locations

2. Sensor Fusion for Real-World Accuracy

We combined:

• accelerometer

• magnetometer

• GPS

To compute:

• direction (compass)

• distance (Haversine formula)

This enables:

• real-time proximity detection

• accurate navigation without maps

3. Offline-First Data Layer

Instead of relying on APIs:

• all monument data is stored locally

• seeded via JSON into a Room database

Result:

• zero network dependency

• instant load times

• predictable performance

4. Hybrid AI + Voice System

We built a dual-mode system.

Online:

• queries go to Google Gemini

• rich contextual responses

Offline:

• keyword-based intent detection

• local knowledge base

Pipeline:

Speech → Intent Detection → Response → TTS

5. Adaptive Intelligence (Network-Aware System)

This is critical for real-world deployment.

Engineering Challenges We Solved

AR Stability Outdoors

Problems:

• lighting variation

• tracking loss

• GPS drift

Solution:

• fallback to sensor-based positioning

• smoothing via filtering

Context Without Internet

Problems:

• no API fallback

• no dynamic fetch

Solution:

• local structured knowledge base

• lightweight intent matching

UX Without Friction

Problems:

• onboarding complexity

• heavy UI

Solution:

• clean Liquid Glass interface

• minimal interaction steps

Impact Beyond a Demo

This isn’t just an AR app.

It’s a cultural infrastructure layer.

Cultural Impact

• preserves local history

• improves engagement

• enables storytelling at scale

Economic Impact

• increases tourism retention

• reduces dependency on guides

• improves visitor satisfaction

Social Impact

• accessible education

• inclusive exploration

• global cultural exposure

Why Samsung PRISM Recognized It

1. Real-world usability

Not a lab demo. Works in messy environments.

2. Mobile-first engineering

Optimized for constraints, not ideal conditions.

3. Scalable architecture

Extensible to:

• cities

• museums

• heritage sites

4. Future metaverse relevance

Bridges:

• physical world

• digital overlays

What Makes This Different

Most AR apps:

• depend on internet

• focus only on visuals

• ignore system design

This system:

• works offline

• integrates AI + voice + AR

• is engineered for real deployment

Key Insight

AR is not the innovation.
Making AR reliable, accessible, and meaningful is.

What We Learned

• real-world constraints matter more than features

• offline capability is a competitive advantage

• UX simplicity beats technical complexity

• systems thinking wins hackathons

Future Roadmap

• 3D interactive models

• occlusion handling

• multilingual support

• large-scale deployment

Conclusion

Campus AR Heritage Guide started as a hackathon idea.

But it revealed something deeper:

Technology becomes powerful when it disappears.

When users don’t think about:

• AR

• sensors

• AI

and just experience:

• history

• context

• place

That’s when it works.

Final Thought

This is not about augmented reality.
It’s about augmenting understanding.

Originally published at: https://www.mendhu.tech/blog/building-an-offline-ar-heritage-guide-from-hackathon-idea-to-samsung-prism-finalist

Autonomous AI agents are no longer experimental. They are now deployed across production systems with access to APIs, databases, internal tools, and critical infrastructure. What changed between 2024 and 2026 is not capability alone, but agency. Agents now:

• take actions without human approval

• chain tools across systems

• operate continuously at machine speed
  Security, however, has not kept pace. This article consolidates research across industry and academia to map the real threat landscape of agentic AI systems.

Most current defenses do not apply.
And most threats remain unsolved.

GITHUB LINK: https://github.com/Atharva-Mendhulkar/AVARA

The Core Problem: Security Assumptions Are Broken

Traditional systems assume:

• humans make decisions

• APIs are deterministic

• permissions are static
  AI agents violate all three. They:

• interpret natural language from untrusted sources

• execute multi-step actions across systems

• escalate privileges through reasoning
  This creates an entirely new attack surface.

Why Existing Security Controls Fail

Traditional controls like:

• firewalls

• RBAC

• IAM
  were designed for static systems. AI agents operate via:

• implicit intent

• probabilistic reasoning

• dynamic tool orchestration

Attacks no longer target endpoints.
They target reasoning pathways.

1. Indirect Prompt Injection: The Invisible Layer

What changed

Direct prompt injection is increasingly mitigated. The real threat is:

Indirect prompt injection Where malicious instructions are hidden inside:

• documents

• emails

• APIs

• tool metadata

Embedding-Level Poisoning (RAG Systems)

In RAG pipelines:

• data is embedded

• retrieved via similarity

• trusted implicitly
  Attackers embed hidden instructions inside vector space.

Impact

• One poisoned document affects multiple queries

• Payload survives vectorization

• No visible attack surface

There is no antivirus for vector databases.

Why this is unsolved

• embeddings are treated as math

• no detection systems exist

• retrieval bypasses access control

2. Multi-Agent Jailbreaks

Instead of breaking rules, attackers bypass them structurally.

Example

Instead of:

“Write malware” Attack splits into:

• Agent A → networking

• Agent B → packet structure

• Agent C → combine
  Each agent is safe individually. Together:

unsafe output emerges.

Core issue

• no inter-agent validation

• no system-level safety

• no reasoning traceability

3. MCP Supply Chain Attacks

MCP (Model Context Protocol) acts like a plugin system. But:

• no vetting

• metadata is trusted

Example

Tool:

add numbers

Hidden:

read ~/.ssh/id_rsa

Why this is dangerous

• models execute hidden instructions

• no permission isolation

• no metadata scanning

4. Excessive Agency

Agents are designed to act. But:

No system enforces when they should stop.

Zero-click exploitation

1. agent reads malicious input

2. interprets as instruction

3. executes autonomously
   No user interaction required.

Root cause

• over-permissioned agents

• no confirmation workflows

• implicit trust

5. RAG as a Privilege Bypass

RAG systems:

• unify data

• ignore original permissions
  Result:

access control is bypassed via similarity search

6. Hallucination Cascades

Multi-agent systems amplify errors.

Chain:

• Agent A → wrong data

• Agent B → validates

• Agent C → builds

• Agent D → decides

error becomes indistinguishable from truth

7. Autonomous Data Exfiltration

Agents can:

• read

• process

• transmit
  without oversight.

New reality

Exfiltration is:

autonomous, contextual, invisible

8. Model Poisoning

Small poisoned samples can create:

• hidden triggers

• persistent backdoors

9. Context Manipulation

Attackers:

• saturate context

• push constraints out

Result:

• degraded reasoning

• constraint loss

10. AI-Accelerated Attacks

AI reduces:

• exploit development time

• attack cost
  From months → hours.

Financial Impact

• Avg breach: $4.88M

• US: $10.22M

• Shadow AI: +$670K

The Core Insight

AI security is not a model problem.
It is a runtime governance problem.

AVARA: A Runtime Security Layer

AVARA introduces:

• intent validation

• tool control

• RAG filtering

• audit logs
  Architecture:

Agent → AVARA → Tools / APIs / Models

Conclusion

AI agents introduce:

• new attack surfaces

• new failure modes

• new risks

The real question is not:
“Can we align the model?” It is:
“Who controls the agent when it acts?”

Originally published at: https://www.mendhu.tech/blog/ai-agent-security-threats-the-complete-landscape-real-risks-and-why-most-defenses-fail

Internal Links

• Read: Building an Offline AR Heritage Guide