- Model Context Protocol
- AI Agent Security
MCP Server Penetration Testing in India
A practical guide for engineering, security and compliance teams, with India's regulatory context and the global standards that apply
MCP Turned AI Assistants into Operators. Test It Like One.
The Model Context Protocol (MCP) lets AI assistants and agents read files, query databases, call APIs and trigger workflows. Every MCP server is therefore a privileged interface that an AI model, and anyone who can influence what that model reads, can drive.
This guide explains the main MCP risks, how an MCP server penetration test works, what Indian regulations and global standards expect, and how to prepare. ISECURION is a CREST-accredited company for penetration testing and a CERT-In empanelled auditing organisation. Our MCP server pentesting service turns these ideas into a scoped engagement.
In this guide:
- What MCP is and why it changes your attack surface
- Eight MCP risks to test for, with real-world incidents
- How an MCP pentest differs from VAPT and LLM red teaming
- CERT-In, DPDP, RBI, SEBI and IRDAI context for India
- Global standards, markets and a pre-test hardening checklist
Request an MCP Pentest Quote
Tell us what you need. We reply within one working day.
What Is MCP, and Why Does It Change Your Attack Surface?
The Model Context Protocol is an open standard that lets AI applications connect to external tools and data in a consistent way. A host such as an AI assistant, IDE or agent runs an MCP client, which talks to one or more MCP servers. Each server exposes tools the model can call, resources it can read and prompts it can reuse. Servers run either locally over stdio or remotely over HTTP.
That design is convenient, and it moves the security boundary. In a normal application, your code decides which function runs. In an MCP workflow, a language model decides which tool to call, and with which arguments, based on text it has read. Some of that text is written by attackers: a web page, a support ticket, a shared document, an email, or even the description of a tool.
Three properties that make MCP servers high-value targets
- They hold credentials and reach. Servers connect to CRMs, code repositories, databases, cloud consoles and payment systems using real tokens.
- They trust model output. Tool arguments come from a model that untrusted content can steer.
- They multiply. A typical agent connects several servers, so one weak or malicious server can influence how the model uses the others.
Why MCP Security Matters Now
India has one of the world's largest developer populations, a fast-growing base of AI-first startups and thousands of global capability centres building internal AI tooling for parent companies abroad. MCP makes it cheap to connect all of that to real systems. It also makes it cheap to connect them badly.
Customer and auditor scrutiny
Enterprise buyers, especially in the US, UK and EU, now add AI and agent questions to security questionnaires. An independent MCP pentest report answers them with evidence instead of assurances.
Regulatory clocks
CERT-In's six-hour reporting rule already applies to Indian entities, and the DPDP Rules begin imposing core duties from May 2027. An agent that leaks personal data through a tool call can trigger obligations under both.
Prototype-to-production drift
Many MCP servers start as templates, community packages or demos. They often reach production with the broad access they were given to make the demo work.
Want the service view instead of the guide? See ISECURION's MCP server pentesting and AI security services for scope, deliverables and FAQs.
Eight MCP Security Risks Every Team Should Test For
These are the failure patterns that show up repeatedly in public research, vulnerability disclosures and the MCP specification's own security guidance.
Prompt injection into tool calls
Instructions hidden in content the model reads, such as a web page, ticket, PDF or email, can steer it into calling tools the user never intended.
Test: indirect injection through every data source a server exposes.
Tool poisoning and rug pulls
Malicious instructions in tool descriptions or schemas are visible to the model but often not to the user. A server can also change its tools after approval.
Test: description, schema and output review, plus change detection.
Weak authentication and authorisation
Authorisation is optional in the MCP specification, so many servers ship with none, with shared API keys, or with checks that cover the server but not each tool.
Test: per-tool and per-user access control.
Token misuse and confused deputy
Accepting tokens issued for other services, passing client tokens through to downstream APIs or mishandling OAuth consent in proxy servers breaks core OAuth boundaries.
Test: audience validation, consent flows, redirect URIs.
Injection in tool code
Tools that build shell commands, SQL queries, file paths or URLs from model-supplied arguments inherit classic flaws: command injection, SQL injection, path traversal and SSRF.
Test: fuzz and exploit every tool argument.
Over-privileged tools and scope creep
A tool that needs read access but holds write or admin rights turns a small mistake into a major incident, and scopes tend to grow over time.
Test: least-privilege review against real permissions.
Cross-tenant and cross-server leakage
Weak session or tenant isolation can expose one customer's data to another, and one server can shadow or manipulate the tools of another.
Test: tenant boundaries and server shadowing.
Supply chain and local tooling
Community servers, npm packages, desktop configs and debugging tools run with developer-level access. Several critical MCP flaws have been in the tooling, not the servers.
Test: dependency, package and configuration review.
We map findings to the OWASP MCP Top 10 (still in beta), the OWASP Top 10 for LLM Applications and the MCP specification's security best practices, so your report speaks the language auditors and customers already use.
What Real-World MCP Incidents Teach Us
MCP is young, yet public disclosures already show how these risks play out. Each case below led to a concrete lesson for how servers should be built and tested.
| When | Case | What happened | Lesson |
|---|---|---|---|
| April 2025 | Tool poisoning research (Invariant Labs) | Researchers showed that instructions hidden in an MCP tool description could make a model leak data, including against a WhatsApp MCP server, while the user saw nothing unusual. | Treat tool metadata as untrusted input |
| June 2025 | MCP Inspector, CVE-2025-49596 (CVSS 9.4) | The debugging tool's proxy accepted unauthenticated requests, allowing code execution on a developer machine through a malicious web page. Fixed in version 0.14.1. | Developer tooling is attack surface |
| July 2025 | mcp-remote, CVE-2025-6514 (CVSS 9.6) | A crafted authorisation URL from an untrusted MCP server could run operating system commands on the client. Versions 0.0.5 to 0.1.15 were affected. | Clients must treat servers as hostile |
| 2025 | Look-alike email MCP package | A malicious npm package posing as a legitimate email service MCP server was reported to leak sensitive data from the people who installed it. | Vet and pin every package |
| 2025 | Asana MCP cross-organisation exposure | A tenant isolation flaw in a SaaS vendor's MCP feature was reported to expose data across organisations. | Test tenant isolation explicitly |
MCP Pentest vs VAPT vs LLM Red Teaming
These three disciplines overlap, and each answers a different question. Knowing which one you need avoids paying for the wrong test.
| Traditional VAPT | LLM red teaming | MCP server pentest | |
|---|---|---|---|
| Core question | Can an attacker break the application or infrastructure? | Can an attacker make the model misbehave? | Can an attacker make the model's tools do harm? |
| Typical targets | Web apps, APIs, networks, cloud | Chatbots, prompts, guardrails | MCP servers, clients, tool schemas, OAuth flows, agent workflows |
| Typical findings | Injection, broken access control, misconfiguration | Jailbreaks, data leakage, unsafe outputs | Tool abuse, token misuse, poisoned tools, cross-tenant exposure, agent-driven privilege escalation |
| Use it when | Always, for the surrounding stack | You care about model behaviour and safety | An AI can call tools or reach your data |
How an MCP Server Penetration Test Works
A clear sequence keeps the work fast and the outputs usable. Each stage produces something your team can act on.
-
Scope and architecture mapping
We inventory hosts, clients, servers, tools, resources, transports and identities, and map where data and credentials flow. We also record which servers are first-party, third-party or community-built.
-
Threat modelling
We identify trust boundaries and the combinations that matter most, such as an agent that can read private data, ingest untrusted content and send data outward. Tools are ranked by blast radius.
-
Authentication, authorisation and transport
We test OAuth flows and token handling against the MCP authorisation model: audience validation, token passthrough, redirect URIs, consent, session management, scope enforcement and stdio launch behaviour.
-
Tool and input testing
We attack every tool the way we would attack an API: command, SQL and template injection, path traversal, SSRF, unsafe deserialisation, schema abuse and missing rate limits.
-
Agent-level adversarial testing
Using realistic agent configurations, we run indirect prompt injection through every data source, tool poisoning, cross-server shadowing, rug-pull simulation, data exfiltration through rendered content and attempts to bypass human approval steps.
-
Reporting, retest and evidence
You receive risk-rated findings with proof-of-concept steps, fixes your developers can apply, a retest to confirm closure and evidence mapped to the OWASP, NIST AI RMF, ISO/IEC 42001 and Indian regulatory requirements you report against.
Ready to scope an MCP pentest?
See scope of work, deliverables, methodology and FAQs on our MCP server pentesting service page.
Regulatory Context for MCP and AI Agents
India has no MCP-specific regulation. The obligations that matter attach to the data, systems and incidents that MCP-connected agents touch. Confirm exact applicability with your legal and compliance advisers.
| Instrument | Why it matters for MCP | What to prepare |
|---|---|---|
| CERT-In Directions (28 April 2022) | Listed cyber incidents must be reported within six hours of noticing them, and logs must be kept within India for a rolling 180 days. An agent-driven data leak can qualify. | Tool-call logging, synchronised clocks, a response plan that names who reports |
| DPDP Act 2023 and DPDP Rules 2025 | Data fiduciaries must apply reasonable security safeguards and notify breaches. The Rules were notified in November 2025, core obligations commence in May 2027, and the earlier SPDI Rules remain in force until then. | A map of what each tool can read, least-privilege access, a breach playbook |
| RBI (banks, NBFCs, payment entities) | IT governance, outsourcing and cyber resilience expectations apply to AI integrations in regulated workflows. RBI FREE-AI Committee report, published in 2025, adds AI-specific governance expectations. | Documented risk assessment and independent testing evidence for each AI integration |
| SEBI CSCRF | SEBI-regulated entities follow the Cybersecurity and Cyber Resilience Framework, which expects testing and audit evidence for systems in scope. | Pentest reports and closure tracking that include AI tooling in scope |
| IRDAI cyber security guidelines | Insurers and intermediaries must govern and test systems that handle policyholder data, including newer AI integrations. | The same evidence pattern: tested, tracked, retested |
Who Needs MCP Testing Most
Fintech and BFSI
Agents connected to customer, payment and lending systems, where a bad tool call carries financial and regulatory impact.
SaaS and product companies
Customer-facing MCP servers and AI features that must prove tenant isolation to enterprise buyers.
Healthtech
Assistants that read patient or claims data, where DPDP safeguards and sector rules apply.
Global capability centres
Internal AI tooling built in India for global parents, which must meet group security standards.
IT services and offshore delivery
Teams using agents and MCP servers on client code and environments, where a compromised developer tool becomes a client breach.
AI-first startups
Agent platforms and MCP marketplaces where security evidence supports funding and enterprise sales.
MCP Security Beyond India: Global Standards and Markets
Most Indian teams ship to global customers, and the strongest MCP security guidance is global. Align to these references and your test evidence travels with you.
MCP specification
Defines OAuth 2.1-based authorisation for HTTP transports, protected resource metadata, resource indicators and an explicit ban on token passthrough. Authorisation is optional, so verify it is actually implemented. Read the authorisation spec.
OWASP projects
The Top 10 for LLM Applications, the Top 10 for Agentic Applications (December 2025) and the MCP Top 10, which is still beta and works best as a reference list.
NIST AI RMF
Govern, Map, Measure and Manage. MCP test results provide measurement evidence and feed risk treatment decisions.
ISO/IEC 42001
The AI management system standard. Pentest results support risk treatment and monitoring controls. See our ISO 42001 services.
Where Global Customers and Regulators Look
United States
Enterprise security reviews, SOC 2 and NIST AI RMF alignment drive most requests for AI integration evidence. See SOC 2 compliance.
European Union and UK
GDPR, DORA for financial entities and the EU AI Act's risk management expectations make independent testing evidence valuable. See GDPR and DORA.
Singapore
MAS Technology Risk Management Guidelines expect regular testing of critical systems, which covers AI integrations connected to them. See our MAS TRM guide.
UAE and Middle East
Financial and virtual asset regulators expect demonstrable security testing from technology providers. See UAE VASP compliance.
Operating across jurisdictions? We map a single MCP test to several frameworks, so you are not testing twice.
Twelve Controls to Put in Place Before the Pentest
Fixing the basics first makes a pentest more valuable, because testers can spend their time on the subtle chains instead of the obvious gaps.
Identity and access
- Require authentication on every remote server, using OAuth 2.1 with per-client consent
- Validate token audience, and never pass client tokens through to downstream APIs
- Enforce authorisation per tool and per user, not only at the server edge
Tool design
- Give each tool the least privilege it needs, and split read tools from write tools
- Validate and constrain every argument, using allow-lists for paths, hosts and commands
- Require human approval for destructive or high-value actions
Agent and prompt safety
- Treat tool descriptions, schemas and outputs as untrusted input
- Pin and review server versions, and alert when tool definitions change
- Separate servers that read untrusted content from servers that can send data outward
Operations and supply chain
- Log every tool call with user, parameters and result, and send the logs to your SIEM
- Run servers in sandboxes with scoped credentials and no shared secrets
- Keep an inventory of approved servers, and block unknown ones
How to Choose an MCP Pentest Provider
MCP testing is new enough that credentials and method vary widely between vendors. Ask every provider the same six questions.
- Independent accreditation. Is the provider CREST accredited for penetration testing, and CERT-In empanelled for audit work?
- MCP-specific method. Can they show how they test OAuth flows, tool poisoning and cross-server risk, not only a generic web test?
- Manual depth. Are findings human-validated with proof-of-concept evidence, rather than scanner output?
- Safe execution. Will they agree rules of engagement for agents that have live tool access?
- Useful output. Are fixes developer-friendly, is a retest included, and are findings mapped to the frameworks you report against?
- Data handling. Where are your data and test evidence stored, and who can access them?
Frequently Asked Questions: MCP Server Pentesting
1. What is MCP server penetration testing?
2. How is an MCP pentest different from normal VAPT or LLM red teaming?
3. Do Indian companies need an MCP server pentest?
4. Which Indian regulations are relevant to MCP security?
5. Is ISECURION CREST accredited for penetration testing?
6. What does the MCP specification say about security?
7. What are the most common MCP vulnerabilities?
8. What is the OWASP MCP Top 10?
9. How long does an MCP server pentest take and what do I receive?
10. Can you test MCP servers in production?
Ready to Test Your MCP Servers Like an Attacker Would?
CREST accredited for penetration testing. CERT-In empanelled. Talk to us about scoping an MCP server pentest or a full AI security assessment, with evidence built for Indian and global audits.