INDUSTRY FACT FILE · AI, ANALYTICS & AUTOMATION

Cybersecurity FAQs for AI, Analytics and Automation

Forty frequently asked questions and answers in five categories, written for Boards, AI leaders, CTOs, CISOs, data protection and risk teams.

In this document 1.  Category 1: AI risk, governance and strategy 2.  Category 2: Regulation, standards and compliance 3.  Category 3: Building secure AI applications and agents 4.  Category 4: Data, models and supply chain 5.  Category 5: Monitoring, fraud and incident response

Audience: Boards, CEOs, AI and data leaders

Q1. Why does AI need its own security approach?

AI systems read untrusted content, hold or retrieve sensitive data and can take actions, creating attack paths such as prompt injection and excessive agency.

Q2. What are the biggest AI risks?

Prompt injection, data leakage, excessive agency, training-data privacy, model supply chain, deepfake fraud, shadow AI and harmful outputs.

Q3. Which incidents illustrate the risk?

A reported US$25 million deepfake video-call fraud (2024), an airline held liable for its chatbot's advice (2024) and confidential code pasted into a public chatbot (2023).

Q4. What should the Board review?

AI inventory and risk register, test results, incidents, data protection status, agent permissions and governance maturity.

Q5. Who should own AI risk?

A cross-functional group led by an accountable executive, with security, data protection, legal, risk and business owners.

Q6. Do we need an AI inventory?

Yes. You cannot govern what you cannot see; list AI systems, models, data sources and agent permissions.

Q7. How should we prioritise investment?

Test customer-facing and agentic AI first, fix data leakage and permissions, then build governance and monitoring.

Q8. How can Codec Networks help?

Codec Networks provides AI/LLM and agent security testing, data leakage testing, model risk reviews, AI governance and DPDPA programmes.

Audience: Legal, compliance and risk teams

Q9. Does India have an AI law?

Not a standalone AI law. MeitY's India AI Governance Guidelines (November 2025) favour principles and rely on existing laws and sector regulators.

Q10. What are the seven principles?

As published, they include trust, people first, innovation over restraint, fairness and equity, accountability, understandability by design, and safety, resilience and sustainability.

Q11. What changed in the IT Rules in 2026?

An amendment on synthetically generated information came into force on 20 February 2026, requiring intermediaries to label synthetic content and act faster on flagged unlawful content.

Q12. How does DPDP apply to AI?

Personal data used to train, ground or run AI must be processed lawfully with notice, safeguards and rights handling; most obligations take effect in May 2027.

Q13. What about BFSI?

RBI's FREE-AI framework guides responsible AI in financial services; SEBI and IRDAI regulated entities should also follow their regulators' expectations.

Q14. Does the EU AI Act matter for Indian firms?

Yes, if you place AI systems on the EU market or their outputs are used in the EU; obligations are phased.

Q15. What is ISO/IEC 42001?

A certifiable standard for AI management systems, published in December 2023, covering governance, risk and lifecycle controls.

Q16. Can Codec Networks help us comply?

Yes. Codec Networks maps AI controls to India's guidelines, DPDP, sector frameworks and ISO/IEC 42001, and tests them.

Audience: CTOs, AI engineering and product teams

Q17. What is prompt injection?

Inputs that override a model's instructions. Indirect injection hides instructions in content the AI reads, such as emails or web pages.

Q18. How do we defend against prompt injection?

Separate instructions from untrusted content, constrain tools, filter outputs, require approvals for sensitive actions and test regularly.

Q19. How should agents be permissioned?

Give each agent the minimum tools and scopes for its task, require human approval for high-impact actions and log every action.

Q20. How do we secure RAG?

Enforce the requesting user's permissions at retrieval, avoid indexing data users should not see and test for leakage.

Q21. How do we protect system prompts and secrets?

Never place secrets in prompts, keep credentials in vaults and assume system prompts may be revealed.

Q22. What should AI testing cover?

Jailbreaks, direct and indirect injection, leakage, tool abuse, output safety, rate limiting and the surrounding application and APIs.

Q23. How do we evaluate output safety?

Use test suites for harmful, wrong or biased outputs, and monitor real conversations with human review.

Q24. Which frameworks guide AI testing?

OWASP guidance for LLM applications, MITRE ATLAS and the NIST AI Risk Management Framework are widely used references.

Audience: Data science, MLOps and privacy teams

Q25. Can we train on customer data?

Only with a lawful basis, clear notice and contractual rights; redact personal data where possible and test for memorisation.

Q26. What is data poisoning?

Manipulating training or fine-tuning data so the model behaves in attacker-chosen ways.

Q27. How do we vet open-source models?

Use approved registries, verify provenance, scan files, prefer safe formats and load models in sandboxes.

Q28. What is an AI bill of materials?

A record of models, datasets and components in an AI system, supporting supply-chain transparency.

Q29. How do we protect model weights?

Restrict access, encrypt storage, monitor downloads and protect inference APIs against extraction.

Q30. How should vector stores be secured?

Apply access control and encryption, segment by tenant or user, and monitor queries.

Q31. How long should AI data be retained?

Only as long as needed for the purpose, with deletion processes for training data, prompts and logs.

Q32. Should staff use public AI tools?

Use approved enterprise tools with data controls; block or govern unapproved tools to prevent leakage.

Audience: CISOs, SOC, fraud and finance teams

Q33. How do we monitor AI in production?

Log prompts, retrievals, tool calls and outputs; alert on injection patterns, unusual actions and leakage.

Q34. What counts as an AI incident?

Data leakage, unauthorised agent actions, model compromise, harmful outputs at scale or AI-enabled fraud.

Q35. Do AI incidents need reporting?

Cyber incidents must be reported to CERT-In within six hours; personal data breaches will need DPDP intimation once provisions commence.

Q36. How do we counter deepfake fraud?

Verify payment and sensitive instructions out of band, train staff and run deepfake simulations.

Q37. What is denial of wallet?

Abuse that drives up AI compute costs; mitigate with authentication, quotas and monitoring.

Q38. Should we run AI red teaming?

Yes, before launch and periodically, especially for customer-facing and agentic systems.

Q39. How should AI incidents be investigated?

Preserve prompts, retrievals and action logs, identify the injection or misuse path and fix controls.

Q40. How can we start with Codec Networks?

Begin with a scoping call on your AI systems, data and agents; Codec Networks will propose a testing and governance plan.

Contact Codec Networks Pvt. Ltd. New Delhi: 507, 5th Floor, New Delhi House, Barakhamba Road, Connaught Place, New Delhi – 110001 Gurugram: 707, 7th Floor, Silverton Tower, Golf Course Extension Road, Sector 50, Gurugram – 122018, Haryana Landline: 011 430 496 96 | 011 437 522 99   Mobile: +91 99716 76124 | +91 88007 88220 | +91 88003 31124 Email: mktg@codecnetworks.com   Web: www.codecnetworks.com

Note: These FAQs provide general guidance as of October 2026 and do not constitute legal or regulatory advice. Confirm obligations against the official instruments applicable to your organisation.