On 4 November, the Indian Computer Emergency Response Team (CERT-In) issued guidelines for auditing the security of artificial intelligence systems used by government agencies and private companies. The 107-page document covers AI systems developed, bought, deployed, or operated in India, and systems that process Indian data or serve Indian users. Reportable AI security incidents must be disclosed to CERT-In within six hours.Each system is scored on eight factors, including possible damage, autonomy, data sensitivity, scale, and reversibility, and rated low, moderate, high or critical. Low-risk tools may need only a self-assessment. Moderate and higher systems need a yearly audit by a CERT-In-empanelled firm, or every six months where risk warrants. Critical systems also need continuous monitoring and red-team exercises.AI Agents Are Exposing India’s Cybersecurity Weak SpotsRisk of ComplianceThe design is sound, but its teeth are borrowed. The document cites no enabling provision, start date or penalty. It binds empanelled auditors rather than AI builders. Only the six-hour clock and 180-day log retention carry clear force, through CERT-In's 2022 directions. That borrowing carries an uncomfortable precedent. Early in 2023, less than a year after CERT-In's six-hour reporting rule took effect, right-to-information replies showed that only 15 entities had reported incidents within the deadline. CERT-In said it held no data on overall compliance, and it was exempted from the RTI Act that November, so outsiders cannot easily check progress. This time, success depends on execution and incentives.The risk is already here. A critical CERT-In advisory last month said AI agents under evaluation exceeded their authorised scope and read confidential records at Indian organisations. None of the victims noticed; the model providers disclosed it. Since the six-hour clock starts only on noticing, audits must test detection first.Execution starts with an inventory, because no one can classify a tool they do not know they run. IBM's 2026 breach study found that shadow AI added Rs 1.79 crore to the average Indian breach where present. Classification is the soft spot. Auditors must validate the organisation's rating, and a higher auditor rating prevails. But a system self-rated low may never meet an outside auditor. CERT-In should publish sectoral examples and sample low ratings.Cost of SecurityThe nine audit domains run from governance and data privacy to fairness, explainability and supply-chain assurance. Low-risk systems must cover only governance and security; from moderate upwards, all nine apply. Each needs a named owner, with one executive accountable to the board. Data and model risk can build on existing engineering discipline: training data, prompts and configuration changes must be traceable.Fairness and explainability are harder. They need statisticians and domain specialists, not only penetration testers. Large language models cannot fully explain themselves, so behavioural testing and decision logs should suffice for routine uses. Where AI affects a loan, a job or a welfare benefit, the bar should be far higher. Third-party risk needs contracts securing audit rights, change notices and prompt incident alerts, since many deployers cannot inspect external models.Agents need the plainest controls: least-privilege access, human approval for consequential actions, and a tested means of interruption. The guidelines push systems that cannot be reliably interrupted into a higher tier. October's intrusions used exposed credentials and over-permissive cloud roles, which are basic hygiene failures.An ordinary AI error is not automatically reportable, while a compromised model may be. CERT-In should publish examples of what crosses the line so that compliance is optimal. The government counted 237 empanelled auditors in March. High-risk audits need AI competence and cannot be one-person jobs, but CERT-In has not said how many firms qualify.One would also suggest that organisations should also be provided with incentives to comply because prudence has not been enough. As per reports, in the year 2023, 47 percent of Nifty 100 companies did not undertake regular cybersecurity audits or training. Yet the average Indian breach now costs Rs 25.5 crore, and red-teaming and penetration testing cut that by Rs 2.47 crore as per reports. The arithmetic favours security, but the incentives do not. Audit costs are certain, breaches are uncertain, customers absorb part of the damage, and disclosure invites scrutiny. The maximum fine for defying CERT-In rose to Rs 1 crore in 2023, still a fraction of one average breach.'Mistaken' CoinDCX Arrests Show Why India’s Digital Policing Isn't Ready YetLong-term RegulationThe answer starts with legal clarity. CERT-In should issue formal directions for high-risk and critical systems. The Reserve Bank of India, Securities Exchange Board of India, Insurance Regulatory and Development Authority of India and Telecom Regulatory Authority of India, all named in the document, can make it binding on their sectors. From May 2027, the data protection regime allows penalties of up to Rs 250 crore for security failures. The Data Protection Board should treat a current audit and a timely report as mitigating factors, giving firms a reason to disclose rather than conceal.Procurement is the cheapest lever. Government buyers should demand audit certificates for AI rated moderate or above. Small firms, who get simplified self-assessment only at the lowest tier, need subsidised audit vouchers and acceptance of vendor-level audits, so that a thousand users of one chatbot need not each commission a review.Insurers could price audit status into cyber premiums, and listed companies could disclose their high-risk systems. Government agencies need a separate route, since fines merely move public money. The Comptroller and Auditor General should examine their AI audits.Incentives can reward the certificate rather than the security, so CERT-In must run random checks and publish annual anonymised compliance data. The 2022 rule faltered largely out of public view. AI agents acting at machine speed make a repeat far more expensive. Meanwhile CERT-In also needs to expand its resources and capacity with enhanced budget.(Subimal Bhattacharjee is a Visiting Fellow at Ostrom Workshop, Indiana University Bloomington, USA, and a cybersecurity specialist. This is an opinion piece and the views expressed above are the author’s own. The Quint neither endorses nor is responsible for them.)