DPDP Act
Digital Personal Data Protection Act, 2023
India's Digital Personal Data Protection Act, 2023, which governs how organisations collect, process and protect personal data.
AI security is the practice of protecting artificial intelligence systems, including models, training data, AI applications and agents, from attacks, misuse and data leakage, and of governing how employees use AI tools. It covers risks such as prompt injection, data poisoning and shadow AI. The term is also used for applying AI to strengthen cybersecurity.
AI security covers three connected areas:
Securing AI systems: protecting the models, training data, prompts, plugins and infrastructure behind AI applications your organisation builds or buys.
Securing AI use: controlling how employees use tools such as ChatGPT, Gemini, Copilot and Claude, so confidential data, customer details or source code don't leak into external services.
Using AI for security: applying machine learning and generative AI to detect threats, analyse alerts and respond faster. This is often called AI in cybersecurity.
When most people search for AI security today, they mean the first two: keeping AI itself safe. Both meanings are covered on this page.
Generative AI has moved from experiments to everyday business tools in a very short time. Employees paste documents into chatbots, developers connect large language models (LLMs) to internal databases, and AI agents are given permission to send emails, update records or run code.
Each of these creates a new attack surface:
AI tools can see sensitive data, so a misconfigured or unapproved tool can leak it.
AI follows instructions in text, so attackers can hide malicious instructions in emails, web pages or documents the AI reads.
AI output is trusted, so manipulated answers can lead to bad decisions or unsafe code.
Attackers use AI too, to write convincing phishing emails, clone voices for deepfake fraud and find vulnerabilities faster.
Traditional security tools were not designed for these risks, which is why AI security has become its own discipline.
The most widely used reference is the OWASP Top 10 for LLM Applications 2025 (source):
ID | Risk | What it means |
|---|---|---|
LLM01 | Prompt injection | Hidden instructions make the AI ignore its rules |
LLM02 | Sensitive information disclosure | The AI reveals personal data, secrets or confidential content |
LLM03 | Supply chain | Compromised models, datasets or plugins from third parties |
LLM04 | Data and model poisoning | Tampered training data changes how the model behaves |
LLM05 | Improper output handling | AI output is passed to other systems without checks, enabling attacks |
LLM06 | Excessive agency | AI agents can take actions far beyond what they need |
LLM07 | System prompt leakage | Attackers extract the hidden instructions behind an AI app |
LLM08 | Vector and embedding weaknesses | Flaws in the knowledge stores used for retrieval (RAG) |
LLM09 | Misinformation | Confident but false output that users act on |
LLM10 | Unbounded consumption | Abuse that drives up cost or knocks the service offline |
Two more risks sit outside the OWASP list but come up in almost every organisation:
Shadow AI: staff using AI tools IT hasn't approved or can't see.
Model theft: attackers copying a proprietary model or extracting it through repeated queries.
For a closer look at the top risk, see prompt injection.
On the defence side, AI helps security teams handle volumes of data no human team could review:
Threat detection: machine learning spots unusual behaviour across endpoints, networks and identities, the core of tools like XDR and UEBA.
Alert triage: AI summarises and ranks alerts so analysts focus on real threats, reducing alert fatigue.
Phishing detection: language models catch convincing AI-written phishing that rule-based filters miss.
Faster response: security automation and AI assistants suggest or carry out containment steps.
Fraud detection: spotting unusual transactions and account takeovers in real time.
AI-driven detection is now a standard part of modern security operations, including AI-powered SIEM and managed SOC services.
Framework | Published by | What it's for |
|---|---|---|
OWASP Top 10 for LLM Applications | OWASP Foundation | The ten most critical security risks in LLM apps, with mitigations |
NIST AI Risk Management Framework (AI RMF) | US National Institute of Standards and Technology | Managing AI risk across four functions: Govern, Map, Measure and Manage |
MITRE ATLAS | MITRE | A knowledge base of real attacker tactics against AI systems, similar to MITRE ATT&CK |
ISO/IEC 42001 | ISO and IEC | A certifiable management system standard for AI, similar to ISO 27001 for information security |
India AI Governance Guidelines | MeitY, Government of India | Principles and recommendations for safe and trusted AI in India |
Most organisations start with the OWASP list for technical controls and NIST AI RMF or ISO/IEC 42001 for governance.
Build an AI inventory: list every AI tool, model and AI-powered feature in use, including ones employees signed up for themselves.
Set an AI acceptable-use policy: state which tools are approved and what data must never be entered into them.
Protect sensitive data: use data loss prevention (DLP) and data classification to stop confidential and personal data reaching external AI services.
Control access: give AI apps, plugins and users only the access they need, following least privilege.
Validate inputs and outputs: filter prompts for injection attempts and check AI output before it reaches other systems or users.
Secure the APIs behind AI: most AI applications run on APIs, so API security controls such as authentication, rate limiting and monitoring apply directly.
Test before and after launch: run AI red teaming to try prompt injection, data extraction and jailbreaks against your own AI apps.
Monitor continuously: log AI usage and watch for unusual prompts, data volumes or agent actions.
Review AI vendors: check where your data is stored, whether it is used for training and what security certifications the provider holds.
AI agents go beyond answering questions: they take actions, such as sending emails, querying databases, booking payments or running code. That makes excessive agency one of the fastest-growing risks.
Give each agent the smallest set of tools and permissions it needs.
Require human approval for high-impact actions such as payments, deletions or external messages.
Treat any content an agent reads (emails, web pages, files) as untrusted, because it can carry hidden instructions.
Log every action an agent takes so it can be reviewed and reversed.
| Aspect | Traditional Cybersecurity | AI Security |
|---|---|---|
| What it protects | Networks, devices, applications, identities, data | AI models, training data, prompts, AI apps, agents and AI usage |
| Typical attacks | Malware, phishing, exploits, credential theft | Prompt injection, data poisoning, model theft, data leakage through AI tools |
| How attacks look | Malicious code or network traffic | Plain text instructions that look like normal input |
| Main controls | Firewalls, EDR, MFA, patching, SIEM | Input/output filtering, AI usage policies, AI red teaming, model and data governance |
Our AI & API Security service puts this into practice for your organisation.
AI & API SecurityDigital Personal Data Protection Act, 2023
India's Digital Personal Data Protection Act, 2023, which governs how organisations collect, process and protect personal data.
An attack that hides instructions in input to an AI model to make it ignore its rules or leak data.
Data Loss Prevention
Tools and policies that stop sensitive data from being leaked, misused or sent outside the organisation.
Using technology to perform security tasks automatically, reducing manual effort and response time.
Protecting application programming interfaces from attacks, misuse and data exposure.
Using AI-generated audio or video that imitates a real person to deceive victims, often to authorise payments.
A list of the ten most critical API security risks, published by the OWASP Foundation.
Employees using AI tools without approval or oversight, risking data leakage.