AI security · Overview

AI security testing & red teaming.

Chatbots, voice agents, AI agents and AI search open new ways to be attacked, and most security teams have never tested them. AI security is HanulSec’s first focus. We attack your AI the way criminals and curious customers will, then help you fix it and prove it.

Scope an AI test See the AI attack map
OWASP LLM and Agentic Top 10India, UAE, Saudi Arabia, worldwideFixed fees, free retests
AI red team · sample● demo
  1. BLOCKED Prompt injection in uploaded PDF
  2. FOUND Agent forwarded invoices to outside address
  3. FLAGGED Voice agent skipped identity check
  4. FIXED AI search now filters by role
  5. PASSED Hinglish jailbreak refused
In plain words

AI security testing finds the ways an AI system can be tricked, misused or made to leak data before attackers do. It covers the model’s hidden instructions, what the AI can see, what it can do and the people around it. HanulSec tests chatbots, voice agents, AI agents and MCP servers, AI search, AI-written code and staff use of AI, and turns the results into fixes and audit-ready evidence.

Services
9 AI security services
Standards
OWASP · MITRE ATLAS · NIST · ISO 42001
Based in
Bengaluru, India
Pricing
Fixed fee per engagement
The AI attack map

Where AI gets attacked, and how we test each part.

Every AI system has the same few weak points: the doors people use to reach it, the model and its hidden rules, and everything it can read or do. Click any part of the map to see how we test it.

Map of where an AI system can be attacked People reach your AI through chatbots, voice agents, AI-built apps and public AI tools. The AI model and its hidden rules connect to company documents, tools and MCP servers, and payment approvals. Attackers can pose as customers, callers or staff. Each part links to the HanulSec test for it. PEOPLE FRONT DOORS THE AI WHAT IT CAN REACH CUSTOMERS chat, app, WhatsApp CALLERS your phone line DEVELOPERS AI coding tools STAFF everyday AI use ATTACKERS can pose as anyone CHATBOTS & COPILOTS tricked into breaking rules LLM RED TEAM → VOICE AGENTS talked into giving data away VOICE RED TEAM → AI-BUILT APPS shipped with open databases APP CHECK → PUBLIC AI TOOLS customer data pasted in SHADOW AI → AI MODEL + HIDDEN RULES system prompt, guardrails COMPANY DOCUMENTS shown to the wrong person DATA-LEAK TEST → TOOLS & MCP SERVERS hijacked by hidden text AGENT & MCP TEST → PAYMENTS & APPROVALS a fake boss on a call DEEPFAKE DRILLS → OUTSIDE AI VENDORS keep what staff paste in RULES YOU ANSWER TO DPDP · CERT-In · ISO 42001 · SOC 2 · EU AI Act COMPLIANCE READINESS → EVERY AI CHANGE new model, prompt or tool AI SECURED MONTHLY →

Scroll sideways to see the whole map, or use the list below.

Services

Nine ways we test your AI.

Start with the system that would hurt most if it went wrong. Most clients begin with their customer-facing chatbot or an AI agent that can take actions.

Coverage

All ten OWASP LLM risks, every time.

The OWASP Top 10 for LLM Applications (2025) is the industry’s list of the ten biggest risks in AI applications. Every HanulSec AI test covers all ten, and agent tests add the OWASP Top 10 for Agentic Applications.

0102030405060708091010/10TESTED

Usually the first to fail

  1. LLM01
    Prompt injection

    Instructions hidden in a message, file, web page or email that make the AI break its rules.

  2. LLM02
    Sensitive information disclosure

    The AI reveals personal data, secrets or another customer’s records.

  3. LLM03
    Supply chain

    Risky models, plug-ins, datasets or libraries the AI depends on.

  4. LLM04
    Data and model poisoning

    Planted data that changes how the model or its knowledge base behaves.

  5. LLM05
    Improper output handling

    AI replies passed to other systems unchecked, opening the door to XSS, SQL or code injection.

  6. LLM06
    Excessive agency

    The AI can take actions or use tools beyond what the task needs.

  7. LLM07
    System prompt leakage

    Hidden instructions, business rules or keys exposed by the AI.

  8. LLM08
    Vector and embedding weaknesses

    Flaws in document search (RAG): missing access checks, poisoned or leaky indexes.

  9. LLM09
    Misinformation

    Confident but false answers that customers or staff act on.

  10. LLM10
    Unbounded consumption

    Abuse that runs up AI bills, slows the service or copies the model.

Know the difference

AI red teaming vs. penetration testing.

You usually need both. A penetration test protects the app around the AI. An AI red team protects the AI itself.

A traditional penetration test

Still essential for servers, apps and networks.

  • Finds bugs in code, servers and logins
  • Follows proven methods such as OWASP and PTES
  • Does not attack the conversation itself
  • Misses prompt injection, jailbreaks and AI data leaks
  • Does not test what an AI agent can be talked into doing

An AI red team

Built for systems that read, talk and act.

  • Attacks prompts, documents, tools and memory
  • Uses the OWASP LLM and Agentic Top 10 and MITRE ATLAS
  • Attacks in the languages your users speak
  • Ranks findings by business harm, not just a score
  • Retests after model and prompt changes
How it works

From first call to free retest.

  1. Day 0

    Free scoping call

    30 minutes with an engineer to agree what to test. A fixed-price proposal follows within one business day.

  2. Day 1

    Permission and NDA

    A mutual NDA and written permission that lists exactly what we may test, and when.

  3. 1 to 4 weeks

    Testing

    Automated attack sets for breadth, then hands-on attacks by our testers for depth. Everything is logged.

  4. End of testing

    Report and walkthrough

    A plain summary for leaders, a technical report with proof and fixes, and a call with your team.

  5. After your fixes

    Free retest

    We rerun every attack that worked. When the retest passes, you get a dated HanulSec AI Tested letter.

Safe by design. We prefer staging environments, agree quiet hours for anything live, never use real customer data in attacks, and stop at once if a system looks unstable.
Standards we map to

Built on the frameworks your auditors use.

OWASP Top 10 for LLM Apps2025 edition
OWASP Top 10 for Agentic Apps2026 edition
MITRE ATLASAttacks on AI systems
NIST AI RMFWith the GenAI profile
ISO/IEC 42001AI management systems
EU AI ActRegulation (EU) 2024/1689
DPDP Act 2023With the DPDP Rules 2025
CERT-InDirections and audit guidelines
SOC 2Trust Services Criteria
ISO/IEC 27001Information security
GDPREU data protection
UAE PDPLFederal Decree-Law 45/2021

Emblems are HanulSec’s own artwork showing which frameworks our work maps to. They are not the owners’ logos and do not mean HanulSec is certified against them.

Learn

Plain-English guides to AI security.

Written by our testers for founders, product owners and security leads.

  1. 6 min readAI securityWhat is LLM red teaming, and does your AI need it?LLM red teaming means attacking your own AI chatbot, copilot or agent the way a real attacker would, before customers or criminals do. Here is what it covers, how it differs from a penetration test, and what a good engagement gives you.
  2. 5 min readAI securityPrompt injection, explained without the jargon.Prompt injection is the number one risk on the OWASP list for AI applications. It happens when text the AI reads is treated as an instruction. Here is how it works, why it is so hard to stop, and what actually reduces the risk.
  3. 5 min readAI securityThe OWASP Top 10 for LLM apps, one risk at a time.The OWASP Top 10 for LLM Applications is the most widely used checklist for AI security. The 2025 edition reflects what went wrong as chatbots became agents. Here is each risk in plain words, with what it looks like in a real product and how to test for it.
  4. 5 min readAI securityMCP security: how AI agents get hijacked.The Model Context Protocol lets AI agents plug into tools and data, from code repositories to inboxes. It also gives attackers new ways in. Here are the attacks seen in the real world so far, why they work, and how to use MCP safely.
  5. 7 min readComplianceCERT-In and AI in 2026, what changed and what to do.In May and June 2026, CERT-In published two documents about AI: a blueprint for defending against AI-assisted attacks, and guidelines for technology vendors on flaws that AI can exploit. Here is what each says, what is binding and what is advice, and a plan to act on them.
  6. 6 min readComplianceThe DPDP Act and your AI chatbot, what to do before May 2027.India’s Digital Personal Data Protection Act never mentions AI, but every chatbot and voice agent that handles customer data is covered by it. Here is what it means for notices, consent, chat transcripts, training data, leaks and children, with a checklist to work through before the main duties apply on 13 May 2027.
  7. 7 min readComplianceEU AI Act Article 50, for chatbots and voice agents.Since 2 August 2026, the EU AI Act has required chatbots and voice agents to say they are AI, and AI-generated content to be marked. These duties can reach Indian companies whose AI serves people in Europe. Here is what Article 50 asks for, who must do it, and how to check that your AI keeps to it under pressure.
Questions

Questions about AI security testing.

Q01What is AI security testing?
AI security testing checks whether an AI system can be tricked, misused or made to leak data. It goes beyond normal security testing by attacking the AI’s instructions, the documents and tools it can reach, and the conversations it has, using the same tricks real attackers use.
Q02What is AI red teaming?
AI red teaming is the hands-on part of AI security testing. Testers play the attacker: they try jailbreaks, prompt injection, data theft and misuse of the AI’s tools, then report what worked and how to fix it. The name comes from military exercises where a “red team” plays the enemy.
Q03Which AI systems can you test?
Customer chatbots and copilots, AI phone and voice agents, AI agents that use tools or MCP servers, AI search over company documents (RAG), apps built with AI coding tools, and the AI tools your staff use. We test systems built on OpenAI, Anthropic, Google, Meta, Mistral and open-source models, on any cloud.
Q04What is the OWASP Top 10 for LLM Applications?
It is a free, widely used list of the ten most serious risks in applications built on large language models, published by the OWASP GenAI Security Project. The 2025 edition covers prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation and unbounded consumption.
Q05How much does AI security testing cost in India?
We quote a fixed fee after a free 30-minute scoping call, so you know the full cost before we start. The price depends on how many AI systems are in scope, whether they can take actions, how many channels and languages they use, and whether you need compliance evidence. We never charge by the number of findings.
Q06How long does an AI security assessment take?
One chatbot usually takes 5 to 10 working days. AI agents with several tools take 2 to 4 weeks. Compliance readiness projects take 4 to 12 weeks. Retesting after your fixes is included.
Q07Do you work outside India?
Yes. We serve India first, then the UAE, Saudi Arabia and clients worldwide. Testing is done remotely from Bengaluru, and on-site visits can be arranged.
Q08Is HanulSec certified?
Our own ISO/IEC 27001 and ISO/IEC 42001 certifications are in progress, and we say so plainly until they are issued. Our testing follows public standards such as OWASP, MITRE ATLAS and NIST, so any expert can check our method.
Next hops

Explore more

Request a quote

Scope it in two minutes.

Tell us what you need protected. A security engineer, not a salesperson, replies with a fixed-price proposal.

Start your quote hello@hanulsec.com
  • Reply within one business day
  • Free 30-minute scoping call
  • Mutual NDA before any detail is shared
  • Fixed price, free retest of fixed findings