OWASP Sounds the Alarm: Vulnerabilities Every Business Must Fix Before AI Eats It Alive

In the past few years, AI systems have quietly moved from research labs into real products used by real people. Models are answering questions, generating code, making decisions, and automating workflows.
Yet one critical question is rarely asked: What happens when these systems are attacked?
Your AI chatbot can be hijacked. And the financial damage can grow faster than you expect. Your training data can be poisoned. This is not a future problem. This is happening right now.

AI is no longer just a productivity tool. It is becoming part of business infrastructure. And once something becomes infrastructure, attackers start treating it like infrastructure too.

Chatbots have taken the world by storm. A hundred million users in just the first two months. We have never seen adoption this fast. LLMs do amazing things, language translation, document synthesis, code generation, decision support.

But every new technology brings people who will try to abuse it. That is why OWASP: the Open Worldwide Application Security Project, released their Top 10 for LLM Applications. OWASP has a long track record of solid security advice. This list is the definitive checklist for anyone building or deploying AI systems.

If your business runs LLMs in production, customer support bots, internal copilots, AI-driven analytics, this is required reading.
Let me walk you through each risk. Simple, scenario-based, with what to do about it.

The Uncomfortable Truth About Building AI Businesses in 2025

The Hard Truth About Building AI Businesses in 2025

This isn’t a dark fantasy. It’s what the data is screaming at us right now.

I’ve spent the past year focused on understanding why brilliant people with great ideas are crashing and burning in the AI space. What I discovered kept me up at night. Not because AI isn’t game-changing (it absolutely is). But because most founders are building these businesses on quicksand.

The Numbers Don’t Lie (But Everyone Ignores Them Anyway)

Research from leading institutions shows that over 90% of AI startups won’t survive beyond 2026. Before you think “that’s just normal startup loss rates,” here’s the kicker: traditional software companies have significantly better odds. AI businesses are failing faster and harder.

MIT published research revealing that nineteen out of twenty enterprise AI pilots never deliver measurable financial returns. These aren’t college kids experimenting. We’re talking about Fortune 500 companies with unlimited budgets, teams of PhDs, and access to the world’s best consultants. And still, 95% fail to show ROI.
In the past two years alone, over $40 billion has been poured into AI initiatives. The return? Disappointingly minimal.
Here’s what really caught my attention: the number of companies abandoning their AI projects has jumped from 17% to 42% in just twelve months. This isn’t a trend settling down. It’s speeding toward a cliff.

The core issue? They’re starting with AI and then searching for problems to solve. The right approach is the complete opposite: start with problems, then decide if AI is even the right solution.

Not every problem needs AI. Not every business should be an AI business. But the hype is so intense that entrepreneurs feel pressured to force AI into everything, no matter if it makes sense.

The Five Deadly Traps Destroying AI Businesses

Navigating the AI Essentials: Unlocking Compliance and Governance in AI

So, here I am, back after months, just to share some essential insights from my recap of 2024. This year, we have all spent an enormous amount of time exploring the edge of AI tools and integrating them into businesses. Still, for me, if I want to summarize the year 2024, undoubtedly it would be AI Compliance and Governance.
You have probably kept track of recent trends in AI development, knowing well that the pace of its progress is exponential. New applications and use cases spring up each day, some of which would be impossible to think of unless just a few years back.

Challenges During AI Advancements

While the progress in AI is remarkable, it hasn’t been without its challenges. Many deployed AI systems have failed to meet expectations, highlighting significant concerns. For instance:

• Inefficient Systems: AI systems in production are not giving the expected outcomes, leading to inefficiencies.
• Misguided Chatbots: The chatbots guide customers and business users into making incorrect decisions.
• Hallucinating Responses: There are incidents of chatbots/agents creating inaccurate responses that spread false information and confusion.
• Biased Results: The use of biased results generated by AI models is a significant cause of concern about ethics and fairness.

The potential of AI cannot be denied; however, companies can suffer massive reputational and financial risks due to early implementation and adaptation.

Why AI Governance is Important

Prompt Engineering: The Key to Unlocking the Full Potential of LLMs

It’s been a while since I last shared my thoughts on this blog, but I’m thrilled to be back to share my thoughts and insights with you all. During my time away, I’ve encountered numerous interesting topics I can’t wait to dive into, but today, I want to focus on something particularly important: Prompt Engineering. This area is crucial for developing Large Language Model (LLM) applications and achieving desired business outcomes. Without further ado, let’s jump into the main topic.

The Significance of Prompt Engineering

In the era of AI and machine learning, particularly when working with LLMs like GPT-4, prompt engineering stands out as a critical area. It’s very important to keep in mind that any LLM is not a magical entity that understands you without explicit guidance. Prompt engineering involves crafting inputs (prompts) to guide the model towards generating accurate and relevant outputs. Also, while writing any prompt, we need to validate what information we are feeding to the LLM, whether we are giving our PII information or providing our DB’s data to get our desired outcome as this is very important to any ERP system or SAAS application. Anyhow, this is a separate topic I will talk about in another blog. This technique is essential for extracting maximum business value from AI models, ensuring they perform tasks as intended and deliver actionable insights.

Types of Prompts

Zero-Shot Learning

Zero-shot learning involves providing the model with a prompt for a task it hasn’t been explicitly trained on, relying solely on its pre-existing knowledge. For example

Prompt: “Generate an SQL query to find all users who registered in the last 30 days.”

Model Output: SELECT * FROM users WHERE registration_date >= NOW() - INTERVAL 30 DAY;

This approach demonstrates the model’s ability to generalize from its extensive training data to new, unseen tasks.

A Beginner’s Guide to BERT Architecture

BERT (Bidirectional Encoders Representational from Transformers) is a pre-trained model for Natural Language Processing(NLP) related tasks. It is introduced by the Google AI team in October 2018 in this paper Pre-training of Deep Bidirectional Transformers for Language Understanding. It is trained on Wikipedia and Corpus dataset, it knows the language and context which is quite decent. It has two versions Base(12 encoders) and Large (24 encoders). It is a specific, large transformer masked language model. BERT is built on top of the multiple clever ideas by the NLP community. Some examples are ELMo, the OpenAI Transformer, and The Transformer.

BERT is a quiet large model that is already trained we just fine-tune it according to specific problems by adding some additional fully connected layers. It is very much expensive to train it from scratch. We have to use the pre-trained model either base(12) or large(24) for our problem. The pre-trained model just understands the language and the context and we do fine-tune to train BERT on a specific task. The overall big-picture of BERT pre-trained and fine-tuning is show in below figure:

Pre-training and Fine-Tuning

As you can see in above figure that pre-trained BERT and how we fine-tune it on to our specific task. For example SQuAD(Stanford Question Answering Dataset) for our question and answering task, NER (Name Entity Recognition) and MLNI (Multi-Genre Natural Language Inference). Two specific tasks on which BERT is fine-tuned are as follows.

• Mask Language Model
• Next Sentence Prediction

How to decompose the time series data?

As in my previous blog, I have discussed white noise in time series data and I said that I will write some blogs on some of the few important concepts in time series forecasting and I would like to share some most important topics which are much needed for better forecasting and analysis. In this blog, I will discuss components of time series data, what is the decomposition of time series data, and how to achieve this in Python.

Basically, decomposition provides a very fine abstract knowledge(generalization) about time series data and it helps to understand the much clear problems during time series forecasting and analysis. Let’s first talk about time series components. There are two main components of abstraction for selecting forecasting methods which are systematic and non-systematic components. Systematic are those components of time series which have consistency in their patterns and can be described directly. Non-systematic are those components of time series that can not be described directly we have to do some pre-processing before describe/model them. A time-series includes three systematic components which include level, trend, and seasonality, and also include one non-systematic component which is noise in time series data.

Let’s just combine the time series components, though a series is a combination of level, trend, seasonality, and noise. For combining these components there are two models, one additive model and another is a multiplicative model. Trend and seasonality are optional because some time series does not have a trend and seasonality through time-series data must have level and noise in it. In the additive model the components are added together as below:
                                 y(t) = Level + Trend + Seasonality + Noise

This model is linear where changes over time are consistently made by the same amount. A linear trend is a straight line. A linear seasonality has the same width and height of cycles. In multiplicative model components are combined together as bellow:
                                 y(t) = Level × Trend × Seasonality × Noise
This model is nonlinear like exponential or quadratic and changes are not constant it can be increased or decrease over time. A nonlinear trend is a curved line and nonlinear seasonality has an increasing/decreasing frequency/amplitude over time.

A gentle intro to White Noise

Currently, I am working on time series forecasting and time series analysis. I’m not going to deep inside the time series analysis though I’m focusing on time series forecasting in my blogs. Before this, I was working on digital image processing, computer vision, machine learning,  deep learning, and some other stuff. From the last couple of days, I start working in time series forecasting and I’m really enjoying this domain because in this domain I convert each time series problem into a supervised learning problem which is really what I like. So basically, I decided to post a few blogs of some important concepts. There are a lot of concepts that are very useful in time series forecasting, natural language processing(NLP) and some other domains as well. I can not post blogs on each and every concept though I will post some fewer topics and will provide a list of remaining concepts in my future blogs.

So, let’s get started with white noise. White noise is a crucially important concept in time series forecasting. If the time series is white noise, it can not be predicted because it doesn’t have any specific patterns, it is just a series of random numbers that can’t be predicted. If the series of forecast does not have white noise, it suggests improvements can be made to the predictive model. So very beginning question is what white noise is?

OpenCV Features of Contours

I am working in Bio-Informatics field. I am in deep of Pathological Myopia Detection and I have a fundus images dataset. One of my tasks is to extract Optic disk from the fundus images which I have done using the UNet model. Now my task was to extract the peripapillary atrophy(PPA) around the Optic disk. For this purpose, I did a lot of research on the Internet and I found many solutions to tackle this problem but the one I love the most is Contours Features. Following is the short intro to Contours features.

First of all, you need to install OpenCV. If you have already installed OpenCV then skip this step. To install OpenCV open your terminal and type the following command:

pip3 install opencv-python

Now, let’s move towards some coding stuff. First of import OpenCV and NumPy and read your desired image and also am resizing my image to 400×400 as follows.

import cv2
import numpy as np
img1 = cv2.imread('myFundImage.jpg',0)
img = cv2.resize(img1, (400,400))

Optic disk mask which is extracted

Now let’s take a look at moments of images. Image moments help you to calculate some features like center of the object, area of the object etc.

Programming with ZeroMQ

Recently I am working with OpenCV and YOLO. When I was just using OpenCV for realtime object detection and recognization there was not good enough acuracy in results after I search a little bit more for acuracy I find two options TencerFlow and YOLOv3. I firstly preffer YOLO algorithm for my desired accuracy and I got good enough accuracy in my results. But I lost the number of frames per second. I have Intel® Core™ i5-2330M CPU @ 2.20GHz × 4 sytem and I was getting 0.6 frames per second(FPS) which is very slow as compare to normal video frames. Then I search a bit more and I realize that YOLO can run on GPU very smoothly. So I got access to GPU and now I face another problem that how could I send my video frames to GPU which is on remote and in response GPU will send back these frames after processing and how could I recieve these frames. Now the main issue is how could sync my local system and remote in no time. I find ZeroMQ solution to solve this issue. ZeroMQ has very simple structure and it is compatible with C/C++, PHP, Java, Python, Lua, Haxe, C#, CL, Delphi, Erlang, F#, Felix, Haskell, Objective-C, Ruby, Ada, Basic, Clojure, Go, Haxe, Node.js, ooc, Perl, and Scala. ZeroMQ (also known as ØMQ, 0MQ, or zmq) looks like an embeddable networking library but acts like a concurrency framework. It gives you sockets that carry atomic messages across various transports like in-process, inter-process, TCP, and multicast. Following figure will show its working in simplest way.

How to install new Magento2 theme manually

Previously I disscussed how to install and configure Magent2 in Linux and now am working in Magento2 development. I faced a problem during installing the new theme in my Magento. I searched a few tutorials on google and some was very confusing so I follow the other documentation. In this blog post I’ll show you that how I install new theme plugin in my Magento2.

I’m going to use Magento in my localhost you can use on your own choice. First of all download your desired theme zip file that you want to install and extract that zip file. Now copy the app and pub folder from extracted theme folder and paste it in your root directory where you install your magento. Since I am running the magento in my localhost so in my case here is the root directory path of Magento.

/var/www/html/magento2

In my case magento2 is root directory of magento. After pasting these two folders open your terminal as root and run few shell commands. Go to your Magento root directory.

~ cd /var/www/html/magento2

If you are using any type of hosting then first connect to your hosting server through ssh. Just go to your magento root directory and follow the same steps below. Firstly run the command to setup upgrade. This command will automatic upgrade all the modules of magento. If any new module is there it will upgraded.