AI Security Assessment
Know what your AI touches, what it can do, and who answers for it.
An assessment of the AI you build and the AI you buy: which personal data trains and reaches it, how it can be attacked or misused, and whether it meets the Act and your regulator's expectations. Tested against OWASP's Top 10 lists for LLM and agentic applications, and mapped to RBI's FREE-AI framework where it applies.
- 40 years with India's regulated enterprises
Trusted by leading enterprises
Why organisations bring their AI to us
Copilots and chatbots are live on real data
Customer and employee data now flows through prompts, uploads and connectors. You need to know what leaves, where it goes and on what terms.
Agents can act in your systems
An agent with tool access can send, change and delete. Its permissions, approvals and failure modes need testing like any privileged user.
Your Board or regulator is asking about AI
RBI's FREE-AI framework recommends Board-approved AI policies, customer transparency and AI incident reporting. Significant Data Fiduciaries also face algorithmic due diligence under the DPDP Rules.
You're buying AI inside your software
AI features in CRM, HR and support tools often arrive with default training rights and overseas processing. Each one is a processor question under the Act.
A complete view of your AI, and how to secure it
Every finding comes with its reference, whether OWASP, MITRE ATLAS, the DPDP Act or your regulator, and a fix your teams can act on.
AI inventory
Every model, AI feature, agent and unsanctioned tool in use, with its data, owner, provider and decision impact.
Lawful basis and notice findings
For each training and inference use of personal data, the ground under the Act and whether notices cover it.
Data leakage findings
What personal data reaches external AI services, through which channels, and on what terms.
Vendor AI terms review
Training rights, retention, sub-processors and data residency in each AI provider's contract.
LLM and agent test report
Rated findings against the OWASP Top 10 for LLM Applications and for Agentic Applications, with proofs and fixes.
Agent permission map
Each agent's tools, identities and approval gates, with over-privileged access flagged.
AI governance gap register
Your policies and controls against FREE-AI, MeitY's AI governance guidelines and your Board's AI policy.
Algorithmic due diligence file
For likely Significant Data Fiduciaries, documented checks that algorithms don't put Data Principals' rights at risk.
Board summary and roadmap
AI risk in business terms, with quick wins and longer-term controls.
From inventory to adversarial testing
Tested against the references that matter for AI in India
Each area carries its technical reference and, where it applies, the DPDP section or regulator expectation beside it.
- Prompt injection
OWASP LLM01 · ASI01
Direct injection, and hidden instructions in documents, web pages and emails the model reads
- Sensitive data disclosure
OWASP LLM02 · DPDP s.8(5)
Personal data leaking through outputs, logs, context windows and retrieval sources
- Excessive agency and tool misuse
OWASP LLM03 · ASI02 · ASI03
Agent permissions, tool scopes, delegated identities and human approval gates
- Supply chain
OWASP LLM04 · ASI04
Models, plugins, tool servers and datasets, and where each came from
- Poisoning and memory
OWASP LLM05 · ASI06
Training data, retrieval sources and agent memory that can be manipulated
- Output handling
OWASP LLM10
Model output reaching other systems without checks
- Adversarial techniques
MITRE ATLAS
Coverage of known attack techniques against AI systems
- Lawful basis for AI
DPDP s.4, s.6, s.7
The ground and notice for each training and inference purpose
- AI providers
DPDP s.8(2), s.16
Contract terms, training rights, retention and where data is processed
- Algorithmic due diligence
DPDP s.10 and Rules
Documented checks for likely Significant Data Fiduciaries
- Financial-sector AI
RBI FREE-AI
Board policy, customer transparency, explainability and incident reporting
AI risk, read through your sector's rules
- Credit and underwriting models checked for explainability and bias, as FREE-AI expects.
- Customers told when they're dealing with AI, with a route to challenge its decisions.
- Third-party AI inside outsourced services assessed under RBI's outsourcing expectations.
AI findings in the same risk picture
The assessment feeds DataSurity, so AI risk sits with the rest of your privacy and security posture.
- Data Journey MappingEach AI use is recorded as a processing activity with its purpose, data, lawful basis and provider.
- Assessment & Compliance ReportingAI findings join the same register and Board report, with owners and retest status.
- Consent ManagementNotices and consent are updated where AI introduces a new purpose.
The inventory, findings and roadmap stay live on DataSurity, so each new AI tool is added to a picture that already exists.
Trusted by leaders across industries
“We knew patient data sat in our hospital information system. We didn't know how much had spread into lab exports, scanned reports and shared drives until DataSurity's assessment showed us. The team understood hospital realities, from paediatric records to staff data, and gave us a plan we could actually run. Implementation is now moving ward by ward, with consent and rights handled in one place.”
Rakesh G
Head - Compliance, KDSG Hospitals (350-bed multispecialty hospital)
Frequently asked questions
Does the DPDP Act apply to AI?
Fully, even though it doesn't mention AI by name. Training data and prompts containing personal data are processing, so they need a lawful ground and a notice. AI providers are processors. Significant Data Fiduciaries must also carry out due diligence that their algorithms don't put people's rights at risk.
We only use AI through vendors. Is this relevant to us?
Yes, often more so. AI features inside SaaS tools can receive customer data through prompts and connectors, sometimes with training rights granted by default and processing outside India. We inventory those uses, test what actually leaves, and review the terms.
Do you test on our production systems or data?
No. Adversarial testing runs in an environment you authorise, with synthetic data. Inventory and lawful-basis work uses documents, logs and interviews. No client data is sent to any external AI service during the assessment.
Is RBI's FREE-AI framework binding on us?
FREE-AI is a committee report RBI released in August 2025, with seven principles and 26 recommendations. It isn't a binding direction on its own, but it signals what RBI expects of regulated entities: Board-approved AI policies, transparency, explainability and incident reporting. Aligning now avoids rework when formal guidance follows.
How is this different from the Cybersecurity Assessment?
The Cybersecurity Assessment tests your applications, infrastructure and controls. This one covers what's specific to AI: data flowing into models, prompt and agent attacks, provider terms and AI governance. Many organisations run both, and the findings land in one register.
How do you use AI in the assessment itself?
Automated attack libraries help us cover the OWASP categories at breadth. Senior testers design the scenarios, validate every finding by hand and judge severity. No client data is used to train anything, and none leaves the agreed test environment.