How Should Indian Enterprises Prepare for AI-Assisted Cyberattacks Under CERT-In’s 2026 Security Guidance?
Almost five months after India's national cyber agency published its blueprint on AI-assisted exploitation, the 12 hour patching figure is still the line that separates prepared companies from the rest. NS3TechSolutions reports on what the document says, what it leaves open and what IT leaders should change first.
14 October 2026 | NS3TechSolutions Editorial
India's national cyber agency has told enterprises to treat a known exploited flaw on an internet facing system as a 12 hour problem. That is the sentence IT teams remember from CERT-In's May 2026 blueprint, and it sits awkwardly beside the two to four week patch cycles that security vendors say many enterprises still run.
The number has travelled further than the document. Some outlets said CERT-In had mandated 12 hour patching. The text says something narrower and more useful: contain, patch or remove exposure "within 12 hours where feasible." The gap between those two readings decides how a company should prepare for AI-assisted cyberattacks, and it is where most of the planning mistakes will be made.
Key points
CERT-In's blueprint is Version 1.0, dated 25 May 2026, 38 pages long. It is guidance. The document states no legal status, and it says organisations are "encouraged" to apply it "in a risk-informed manner."
The remediation tiers run from 12 hours for known exploited flaws on internet facing and crown jewel systems to 5 days for high severity flaws. Where no patch exists, isolation, access restriction, WAF protection or disabling a feature counts as a response.
The blueprint goes beyond patching. It asks for AI system inventories, shadow AI monitoring, controls on agentic AI, executive verification against deepfakes and a Bill of Materials for software and AI components.
It restates the existing 6 hour duty to report cyber incidents to CERT-In, and sets a 7, 30 and 60 day roadmap.
IBM's 2026 India data shows the cost of falling behind: breaches averaged INR 25.5 crore, and INR 31.6 crore where AI and security automation were absent.
What did CERT-In actually publish?
The title is "Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure." It carries no document number. Its argument, in CERT-In's words as reported by MediaNama, is that "exploitation timelines are reducing significantly," that attacks are "becoming increasingly autonomous," and that "traditional static security approaches would become insufficient."
It describes five threat families. AI speeds up reconnaissance and exploit development. AI writes more convincing phishing and powers deepfake voice and video fraud. AI generates malware that changes shape between attacks. AI systems themselves can be attacked through prompt injection, data poisoning and model theft. And all of this lands on vital sectors, which the document lists as government, finance, telecom, digital public infrastructure, healthcare, energy and transport.
The remediation expectations in Section 9 are the part that changes calendars:
Known exploited vulnerability on an internet facing or crown jewel system: immediate containment, and patch, mitigate or remove exposure within 12 hours where feasible
Critical externally exposed vulnerability: patch, mitigate or remove exposure within 1 day
Known exploited vulnerability on an internal system: patch or mitigate within 1 day unless compensating controls are in place and documented
Critical internal vulnerability on a high value system: within 3 days
High severity vulnerability: within 5 days, based on risk prioritisation
Section 12 then lists 16 areas of AI governance, from oversight and inventory to prompt injection testing, third party AI provider risk, model integrity and the governance of autonomous agents. The blueprint also points to a CERT-In AI Cyber Defence Center, with a separate contact address for AI compliance matters, while incident reporting stays on the existing channel.
This is not CERT-In's first move on the subject. A late April 2026 advisory, reported by Business Standard, warned about frontier models such as Anthropic's Claude Mythos, citing their dual use nature, and told companies to shrink internet facing attack surfaces, segment networks, harden or replace outdated remote access systems such as older VPNs, and restrict outbound traffic to known AI services. The May blueprint turns that alert into a structured programme.
Is the blueprint mandatory?
No, not in the sense of a rule with penalties. Khaitan & Co's analysis says the blueprint is not backed by a statutory rule making provision and is not binding, but advises reading it as a signal of where regulatory expectations are heading rather than as purely advisory. JSA's commentary describes it as guidance rather than binding regulation, issued under CERT-In's powers under Section 70B of the IT Act, 2000. The Register also read it as recommendations.
What is binding is older. The CERT-In Directions of 28 April 2022 already require specified incidents to be reported within six hours. The blueprint repeats that requirement, so a company that cannot detect and escalate within that window is already exposed, whatever it decides about the rest.
Our reading: treat the document as the benchmark an auditor, a bank's vendor risk team or a regulator will reach for after an incident. Khaitan notes that if the timelines were ever codified they would require significant workflow improvements. Companies that start now will not be building under deadline.
Why 12 hours? What the exploit data shows
The number looks arbitrary until you set it next to the attacker side of the ledger.
Mandiant's M-Trends 2026, as summarised by ComplexDiscovery, puts the estimated mean time to exploit newly disclosed vulnerabilities at minus seven days. In plain terms, on average the exploit appears about a week before a patch exists. Exploitation of internet facing systems was the leading initial infection vector for a sixth year running, at 32% of cases where the entry point was identified. Attackers handed off access in as little as 22 seconds, against more than eight hours in 2022. These are Mandiant estimates, and the summary we reviewed does not spell out the method behind the minus seven figure.
AI changes the cost side. A research system called CVE-Genie reproduced working exploits for about 51% of 2024 and 2025 CVEs at roughly USD 2.77 each, according to a preprint cited in a Cloud Security Alliance research note. That is a research prototype, not an attacker toolkit, but it shows the direction.
Defenders move slower. Verizon's 2025 Data Breach Investigations Report, as summarised by CNiC Solutions, found only 54% of vulnerable edge devices were fully remediated, with a median of 32 days. For comparison, The Register notes that US federal deadlines under CISA's Known Exploited Vulnerabilities catalogue run from a few days for the most serious bugs to two or three weeks.
Put together, the 12 hour figure reflects what attackers can do more than what most defenders can do today. The Cloud Security Alliance note makes the same point and adds that most enterprises, and nearly all small ones, cannot reliably meet it without automation and monitoring. That last claim rests on thin support and should be read as the authors' inference.
How ready are Indian organisations?
India's baseline is busy. CERT-In handled over 29.44 lakh incidents in 2025, according to a January 2026 government review published by the Press Information Bureau. Most of that is probing and scanning, but it is the background against which a SOC works.
IBM's 2026 Cost of a Data Breach findings for India, released on 3 August, put the average breach at INR 25.5 crore, up 15.9% from INR 22 crore. IBM found 26% of malicious breaches in India were AI generated. Phishing was the most common way in at 19%, ahead of drive by compromise at 16% and supply chain compromise at 15%.
The preparedness numbers are the sharper story. Only 32% of Indian organisations in the study reported extensive use of AI and security automation, 36% reported limited use and 32% none. Cost tracked usage:
No AI or automation: INR 31.6 crore per breach
Limited use: INR 23.1 crore
Extensive use: INR 21.3 crore
That is a gap of INR 10.3 crore between the two ends. Shadow AI, meaning staff using AI tools IT has not approved, added an average of INR 1.79 crore per breach where it was present. Offensive security testing, such as red teaming and penetration testing, was the largest single cost reducer at about INR 2.47 crore. IBM's study is a Ponemon Institute survey of 602 organisations worldwide, and the release summary we reviewed does not give the India sample size, so treat the India splits as indicative.
Intent is lagging too. Among organisations planning post breach investments, 67% named incident response plans and testing, but only 39% named AI security and governance tools. Globally, IBM reported an average breach cost of USD 4.99 million and said roughly one in four malicious breaches was AI enabled, at about USD 6 million each.
In finance, a June 2026 RBI survey summarised by TechTimes found banks and NBFCs ranking AI-enabled cyberattacks first among near term threats, and 93% of 43 surveyed institutions relying on outside vendors for core cybersecurity. We have not seen the underlying RBI table, so verify before quoting.
The part most coverage skipped: AI systems as an attack surface
Headlines about patching crowded out half the document. Section 12 treats the company's own AI as something to defend.
Shadow AI and public AI platforms
The blueprint asks for an inventory of AI systems and integrations, including unsanctioned ones, and for limits on uploading sensitive data to public AI platforms. Khaitan's summary adds AI usage policies, approval of AI integrations and monitoring of AI APIs. IBM's INR 1.79 crore shadow AI premium suggests this is a cost item, not just a policy item.
Practically, the first obstacle is discovery. Most firms can list their servers. Few can list every browser plugin, copilot, no code agent and API key that now touches company data. Network egress logs, identity provider records and expense reports are the usual places to start.
Agentic AI
For autonomous agents, the blueprint asks for defined operational boundaries and permissions, continuous monitoring and audit logging, human oversight of critical decisions and emergency shutdown mechanisms. It also asks for an AI Bill of Materials alongside software, quantum and cryptographic equivalents, building on CERT-In's July 2025 technical guidelines on bills of materials.
Our reading: an agent with a service account and broad access is an identity risk before it is an AI risk. Treat it like a privileged non human user, and the existing controls start to apply.
Deepfakes and executive impersonation
The blueprint calls for executive verification procedures, monitoring for executive impersonation, escalation paths that include law enforcement, and staff training on AI-assisted phishing and deepfakes. It does not prescribe a technique, so each firm has to choose its own.
The consumer evidence shows why. McAfee's 2026 State of the Scamiverse, a November 2025 online survey across seven countries, found one in five Indians had met a voice clone scam, 51% reported losing money to a scam and the average loss was INR 93,915. The India sample size is not given, and this is consumer data, not enterprise data. But the same voice and video tools are what a fraudster uses on a finance team.
A workable rule is simple. Any payment change, credential reset or urgent instruction that arrives by voice, video or message gets confirmed through a second, pre registered channel, regardless of how senior the apparent sender is.
Where the blueprint stops short
MediaNama's reading identifies four gaps, which we reproduce as that publication's analysis:
No sector specific guidance, despite different obligations for healthcare, finance and telecom
AI governance measures sit with operational IT teams without a clear link to board level oversight (JSA's commentary, by contrast, reads the blueprint as calling for board and executive oversight, so readers differ)
Silence on open source AI models, which have no vendor, contract or formal supply chain to assess
The July 2025 audit guidelines require annual third party audits that cover AI system security, but the AI measures here are voluntary and nothing links the two
A fifth gap is practical. The document gives no guidance on how to meet 12 hours in operational technology, legacy or vendor locked systems. It accepts temporary mitigation, which helps, but leaves the design to the reader.
How it fits with existing Indian rules
Three clocks already run in parallel, and the blueprint adds a fourth.
CERT-In Directions 2022: report specified incidents within six hours of noticing them, keep ICT logs for 180 days and synchronise clocks to NIC or NPL sources
DPDP Rules 2025: notified 13 November 2025, with a detailed breach report to the Data Protection Board within 72 hours and a minimum one year log retention. Advisers cite penalties of up to INR 200 crore, and one report puts full enforceability at 13 May 2027
RBI Directions of 31 July 2026: summaries describe a 24/7 cyber SOC and six hour reporting for regulated banks. At least one adviser summary treats NBFCs differently from RBI's own listing, so confirm your category against the RBI text
The blueprint: a 12 hour containment expectation for the most exposed flaws
Our reading: the common thread is evidence. In each case the question after an incident will be what you knew, when you knew it and what you did. That is a logging and SOC question as much as a patching one.
What should enterprises do? A 7, 30 and 60 day plan
The blueprint's own roadmap is a sensible spine. We have added the questions we would ask at each stage.
Days 0 to 7: immediate risk reduction.
Name an owner for the programme. List every internet facing asset and the crown jewel systems behind them. Turn on MFA for all critical and privileged access. Run vulnerability assessments and apply fixes or mitigations for flaws on CISA's and other known exploited lists. Confirm that logging is on and that an incident can be escalated inside the six hour window. Begin training on AI-assisted phishing and deepfakes. Test: if a known exploited flaw was announced at 6 pm tonight, who would see it, who could approve a change and who could isolate the system?
Days 8 to 30: operational strengthening.
Integrate endpoint, cloud, identity and network telemetry into the SOC. Stand up continuous attack surface monitoring. Create the AI inventory and a shadow AI policy. Review cloud and API exposure. Run a tabletop exercise on a ransomware case and a backup restoration test. Test: can you restore a critical system from backup in the time the business says it can tolerate?
Days 31 to 60: advanced resilience.
Run red team or purple team exercises. Test AI applications for prompt injection and model integrity. Expand automation for triage and response, with human approval for high impact actions. Decide which parts of the estate cannot meet 12 hours and document the compensating controls. Test: have you proved a mitigation works, or only recorded that it was applied?
The practical design choice is which of the 12 hour actions to automate. Isolation rules, WAF virtual patches and account disablement can be pre approved as standing changes, so the clock is spent on containment, not committee. That is our recommendation, consistent with the blueprint's acceptance of mitigation in place of a full patch.
What the experts say
Dray Agha of Huntress said, as reported by The Register, that the guidance is sound because it does not demand a full patch in 12 hours, which turns the deadline into an achievable containment strategy. He said Huntress has seen vulnerabilities exploited within a few hours even before AI-assisted attacks became common, and argued the guidance pushes organisations towards continuous defence in which business functions beyond IT must take part. The Register's own reporting notes that practitioners it spoke to thought 12 hours too short to fully test and deploy a patch.
Qualys, in a June blog post, argues that change advisory boards and two to four week patch cycles are structurally incompatible with the 12 hour bar. That post is written by a vendor that sells continuous risk tooling, and several statistics in it carry no source, so we cite it only for the argument, not the figures.
IBM India's Gaurav Agarwal said that "AI with agentic capabilities must be embedded across the full security lifecycle." IBM also reported that 78% of 456 surveyed organisations were aware of frontier AI models such as Mythos.
What happens next
These are forecasts and our own reading, not facts.
Expect the blueprint to be cited in vendor questionnaires and audit scopes well before it appears in any binding rule. The July 2025 audit guidelines already require annual third party audits, and auditors tend to adopt reference documents quickly. Sector regulators are likely to set their own timelines for banks, insurers, telecom and healthcare, as the missing sector guidance becomes harder to ignore. And as DPDP obligations reach full force around May 2027, breach evidence will matter more because penalties attach to notification failures.
Attackers will not wait. CERT-In itself expects exploitation timelines to keep shrinking. The sensible planning assumption is that 12 hours is a staging post, not a ceiling.
How NS3TechSolutions Helps Prevent AI-Assisted Cyberattacks
Meeting these expectations is mostly an infrastructure and operations problem before it is an AI problem. A SOC can only act on telemetry it receives, a containment order can only be executed if the network is segmented and the firewall policy is under control, and an asset list can only be trusted if someone has verified it.
That is where NS3TechSolutions works: network and security architecture, firewall and perimeter design, endpoint and cloud security, and managed SOC and NOC services. In practice, we would help a team do three things. First, assess whether the network, endpoint, identity and cloud estate produces the logs and exposure data a continuous programme needs. Second, design the standing mitigations (segmentation, access restriction, virtual patching) that let a team contain a flaw inside 12 hours without waiting for a vendor patch. Third, scope a managed SOC or NOC arrangement against the real asset list, with written authority to contain and evidence ready for the six hour report.
Practical checklist
Save this for your next audit or board update.
Name an executive owner and write the 12 hour, 1 day, 3 day and 5 day tiers into your vulnerability policy
Maintain a live list of internet facing assets and crown jewel systems
Enable MFA on all privileged and remote access
Pre approve containment actions (isolate host, disable account, block IP, virtual patch) so they need no committee at night
Define which systems cannot meet 12 hours and document compensating controls
Confirm detection and escalation can meet the six hour CERT-In reporting window
Keep logs for at least 180 days, and one year where DPDP applies, with time synchronised
Create an AI inventory, including unsanctioned tools, and a policy on uploading data to public AI platforms
Treat AI agents as privileged identities with limits, logging and a shutdown switch
Start an AI Bill of Materials alongside your software one
Add a second channel verification step for payments and credential changes requested by voice, video or message
Run a tabletop exercise, a backup restoration test and a red team exercise this year
FAQ
Q. Is CERT-In's AI blueprint mandatory for private companies?
A. No. The document states no legal status and says organisations are encouraged to implement it in a risk-informed manner. Law firm commentary treats it as guidance that signals regulatory direction. The six hour incident reporting duty it repeats is binding under the 2022 Directions.
Q. Does CERT-In require every vulnerability to be patched in 12 hours?
A. No. The 12-hour expectation applies to known exploited vulnerabilities on internet facing and crown jewel systems, and it allows containment, mitigation or removal of exposure where a full patch is not feasible. Other findings have 1, 3 or 5 day tiers.
Q. Can a company meet these timelines without a 24/7 SOC?
A. It is difficult. A 12 hour window can open and close overnight or over a weekend. Companies without in house night cover typically pair internal decision makers with a managed SOC that holds written authority to contain.
Q. What counts as a compensating control?
A. The blueprint lists isolation, access restriction, WAF or API protection, enhanced monitoring and disabling the affected feature. The control should be documented and tested, not just recorded.
Q. What should an AI inventory include?
A. Approved and unapproved AI tools, models, APIs, plugins and agents, what data each touches, who owns it and which third party provides it. The blueprint also expects an AI Bill of Materials.
Q. How does the blueprint relate to DPDP and RBI rules?
A. They are separate instruments. The blueprint is guidance. DPDP imposes a 72 hour breach report and log retention, and RBI's 2026 directions set sector duties for regulated entities. A single incident can trigger several clocks at once.
Q. What is the first thing to do this week?
A. List internet facing assets, confirm MFA on privileged access and run the 6 pm test: who sees a known exploited flaw, who approves action and who can isolate the system.
The number to take to the board
Twelve hours is a measure of decision speed, not patching speed. The company that can name who decides, who acts and what has been pre-approved will meet the intent of CERT-In's blueprint even when a vendor patch is days away. The company that has to convene a meeting will not, whatever its patch dashboard says. With breach costs in India averaging INR 25.5 crore and AI automation separating INR 21.3 crore outcomes from INR 31.6 crore ones, that is a question worth settling before the next advisory arrives, not after.