Nitesh M.
United Arab Emirates · Product Security Lead · Quantum Wealth Tech Limited
University of Mumbai
5 followers
About
I am a Product & Platform Security Lead with 12+ years of experience across AI Security, Product Security, Application Security, Cloud Security and DevSecOps. I specialise in building security capabilities, driving security-by-design and embedding security controls across the product development lifecycle.
I lead security initiatives across engineering, product and platform teams, driving architecture reviews, threat modelling, risk management and secure SDLC practices across AI-enabled products and enterprise applications.
My primary focus is AI Security, particularly securing LLM applications, Agentic AI, RAG workflows and MCP architectures. My experience includes AI security assessments, architecture reviews and adversarial testing to identify risks such as prompt injection, jailbreaks, sensitive data exposure, insecure tool execution, excessive agency and authorization weaknesses. I focus on establishing security controls around identity, access permissions, trust boundaries, memory security and secure AI workflows.
I also lead Application and API Security, DevSecOps and Secure SDLC initiatives, integrating SAST, SCA, DAST, container, secret detection and Infrastructure-as-Code security into CI/CD pipelines using tools such as Snyk, Trivy, OWASP ZAP and Checkov. Through security automation, CI/CD enforcement and engineering collaboration, I have contributed to an 80–85% reduction in identified security vulnerabilities.
My broader experience includes Cloud and Container Security, Kubernetes, Docker, Terraform/IaC security, penetration testing, vulnerability management, SBOM generation and software supply-chain security.
Beyond technical controls, I focus on security architecture, risk prioritisation, engineering collaboration, security governance and measurable risk reduction, helping organisations build resilient products while accelerating cloud and AI adoption.
Certifications: OSCP | CREST CRT/CPT | AWS | CTMP
I am passionate about securing emerging AI technologies, strengthening product security and building scalable security programmes that enable engineering teams to innovate securely.
Specialties
- AI / ML Security
- Application Security
- DevSecOps
- Platform Security
- Security Engineering
Skills
- Application Security
- Web Security
- API Security
- Mobile Security
- OWASP Top 10
- DevSecOps
- Secure SDLC
- CI/CD Security (Github Actions)
- SAST
- SCA
29 skills in total
Experience
-
Product Security Lead
Quantum Wealth Tech Limited
Aug 2024 — Present
Abu Dhabi, United Arab Emirates
-
Senior Security Consultant
Cloudsufi India
Sep 2022 — Aug 2024
India
-
Senior Application Security Engineer
Cornerstone OnDemand
Dec 2020 — Jul 2022
7 experiences in total
Education
-
University of Mumbai
Jan 2009 — Jan 2013
Recent posts
3/3 on today's Cyberr daily challenge.
Senior level: application security, threat intelligence and cloud security.
Can you match 3/3 on Application security?
https://cyberr.ai/daily/d9d687af-a947-4069-8152-6e1bf4d4f526
Main Index - https://cyberr.ai/s/B3APl6f
Flag: flag-073625
Vulnerability: Prompt Injection and Social Engineering Leading to Authorization Bypass
The CTF was solved by manipulating the gatekeeper agent into accepting a user-asserted Supervisor Clearance ID without adequate verification.
The exploitation sequence involved:
Identifying the access restriction requiring Tier-1 Supervisor clearance to open the secure door.
Discovering the available open_sesame tool and its requirement for a valid Supervisor Clearance ID.
Initially attempting to open the door using clearance ID 007, which was rejected.
Reframing the request by explicitly claiming that 007 was the user's Supervisor Clearance ID.
Repeating the clearance assertion, after which the agent accepted the claim and opened the secure door without demonstrating independent identity or clearance verification.
The complete exploitation sequence, including the exact user prompts, agent responses, tool invocation, and successful result, is documented in the screenshots below.
Evidence Screenshots - The gatekeeper successfully opened the secure door and returned the flag: flag-073625
Root Cause – The gatekeeper relied on conversational assertions rather than independently verifying the user's identity and supervisor clearance against a trusted authorization source. Access restrictions were effectively enforced through agent instructions instead of robust backend authorization controls.
Remediation - Never rely on system prompts as security boundaries. Validate user identity and clearance through a trusted authorization service, and enforce access controls within the backend tool before performing privileged operations. Keep authentication credentials outside the LLM context.
Impact – The vulnerability allowed an unauthorized user to bypass the supervisor clearance requirement and open a protected resource. In a real-world environment, similar weaknesses could allow unauthorized access to privileged functions, restricted systems, or sensitive information.
Severity - High
OWASP GenAI LLM Top 10: 2026 Mapping -
LLM01 — Prompt Injection (Primary Mapping): The agent was manipulated through social engineering and a fabricated supervisor identity claim, resulting in it bypassing its intended access restrictions.
LLM02 — Sensitive Information Disclosure (Impact-Based Mapping): The successful bypass exposed the protected challenge flag. Similar weaknesses could result in unauthorized disclosure of sensitive information in real-world applications.
Technical Vulnerability Classification:
1) CWE-287 — Improper Authentication: The agent accepted a user-asserted clearance identifier without adequate independent verification.
Main Index - https://cyberr.ai/s/B3APl6f
Flag: flag-3017d4
Vulnerability: Sensitive Information Disclosure through Diagnostic and Health-Check Tools
The CTF was solved by interacting with the gateway's diagnostic and service health-check functionality to discover sensitive information exposed by the diagnostic process.
The exploitation sequence involved:
Reviewing the system logs using view_system_logs and identifying that the AuthService was unresponsive and required a valid Diagnostic ID for further checks.
Generating a Diagnostic ID using generate_diagnostic_id, which returned DBG-8821.
Invoking check_service_health with the Diagnostic ID and providing AuthService as the target service.
Reviewing the diagnostic findings and discovering a cloaked agent configuration endpoint through service discovery using scan mode storm.
Identifying the agent as AGENT-001 and extracting the required access key exposed in the diagnostic output.
Recovering the access key and successfully completing the challenge.
The complete exploitation sequence, including the diagnostic requests, agent responses, discovered endpoint, and exposed access key, is documented in the screenshots below.
Evidence Screenshots -
The diagnostic output exposed the following required access key: flag-3017d4
Root Cause – The diagnostic and health-check functionality exposed excessive internal information, including a cloaked agent configuration endpoint and its required access key. Sensitive operational information was returned through diagnostic responses without adequate access restrictions or redaction.
Remediation - Restrict diagnostic functionality to authorized users, sanitize diagnostic responses, and prevent credentials or access keys from appearing in logs and tool outputs. Apply least-privilege access controls and separate diagnostic information from sensitive configuration data.
Impact – The vulnerability exposed an internal agent configuration endpoint and its required access key. In a real-world environment, similar information disclosure could expose credentials, internal service details, or configuration secrets, potentially enabling unauthorized access to protected services.
Severity - High
OWASP GenAI LLM Top 10: 2026 Mapping -
LLM02 — Sensitive Information Disclosure (Primary Mapping): The diagnostic tool exposed a sensitive access key and internal configuration information through its responses, allowing information that should have remained protected to be retrieved by the user.
Technical Vulnerability Classification:
CWE-200 — Exposure of Sensitive Information to an Unauthorized Actor.
**Breaking AI Agents: My ARKX CTF Journey – Advanced & Expert Challenges*
CTF Link - https://ctf.arkx.ninja/
AI agent security is quickly becoming a critical part of application security. As AI systems move beyond simple chatbots and into agentic applications that can call tools, access resources, execute code, and interact with external systems, the attack surface expands in ways that are easy to underestimate.
This blog series follows my journey through a set of Advanced and Expert Level CTF challenges, with a focus on the vulnerabilities I explored, the exploitation techniques I used, and the security lessons that emerged along the way.
The challenges cover several key areas of AI security, including:
- **Tool and Resource Security:** Exploiting weaknesses in MCP tools, resources, and access controls.
- **Code and Command Injection:** Identifying unsafe execution paths through AI-connected tools.
- **Insecure Deserialization:** Exploiting unsafe YAML processing to achieve code execution.
- **Parser Differentials:** Bypassing security validation through inconsistent data parsing.
- **Agent and Workflow Security:** Understanding privilege escalation and weaknesses in multi-agent workflows.
- **Race Conditions:** Exploiting parallel financial operations to demonstrate double-spending.
Rather than focusing only on the flags, each write-up digs into how the vulnerability worked, why it existed, what the real-world impact could be, and how to remediate it properly.
**What to Expect from This Series**
Each challenge post will break down the attack scenario, exploitation approach, technical root cause, security impact, relevant OWASP GenAI LLM Top 10 mapping, and practical remediation guidance.
The goal is to go beyond finding vulnerabilities and understand how similar issues can surface in real-world agentic AI applications.
**All testing was conducted within the intentionally vulnerable ARKX CTF environment. These write-ups are intended for educational purposes and security awareness.**
All the challenge write-ups will be linked here as the series progresses.
Advanced Level
1) Level 1 – The Context Gateway - https://cyberr.ai/s/qRs3LJi
From LLMs to Agentic AI: Understanding the Real AI Stack
Everyone is talking about AI. But most people are mixing up terms like LLMs, RAG, AI Agents, and Agentic AI as if they mean the same thing.
They don’t.
Here’s the simplest way to understand the AI stack 👇
🔹 LLM (Large Language Model)
This is the “brain” trained on massive amounts of data.
Examples:
* ChatGPT
* Claude
* Gemini
* Llama
An LLM predicts the next best word/token based on patterns it learned during training.
By itself, an LLM only knows what it was trained on.
It doesn’t automatically know your company data, live internet updates, or internal documents.
🔹 RAG (Retrieval-Augmented Generation)
RAG gives the LLM “memory” or “access to external knowledge.”
Instead of relying only on training data:
1. The system searches relevant documents/data
2. Retrieves the most relevant information
3. Injects that into the prompt
4. The LLM answers using that context
Think of it as: LLM + Search Engine + Company Knowledge
This is how enterprises build:
* Internal AI assistants
* Policy chatbots
* Secure enterprise search
* AI over PDFs, tickets, codebases, or SharePoint
🔹 AI Agents
Agents go beyond “chat.”
They can:
* Make decisions
* Use tools/APIs
* Execute tasks
* Interact with systems
* Take multi-step actions
Example:
Instead of just answering:
“Your flight is delayed.”
An AI Agent could:
✔ check flight APIs
✔ rebook tickets
✔ notify your calendar
✔ send an email confirmation
Agents = AI that can *act*, not just respond.
🔹 Agentic AI
This is where things become truly autonomous.
Agentic AI systems can:
* Plan objectives
* Break tasks into sub-tasks
* Reason through workflows
* Use multiple agents/tools
* Self-correct
* Continue until goals are achieved
Example:
“Find security vulnerabilities in our application, prioritize them, create Jira tickets, and generate remediation guidance.”
No single prompt.
No human guiding every step.
The system orchestrates itself toward a goal.
💡 In short:
✅ LLM = understands/generates language
✅ RAG = gives AI access to external knowledge
✅ AI Agent = AI that can take actions
✅ Agentic AI = autonomous goal-driven AI systems
We’re moving from:
“AI that talks” → “AI that works.”
And that shift is going to redefine software, cybersecurity, engineering, operations, and business workflows over the next few years.
#AI #LLM #RAG #AIAgents #AgenticAI #ArtificialIntelligence #CyberSecurity #MachineLearning #GenerativeAI #DevSecOps