India's national cybersecurity agency has told organisations running internet-facing systems to patch, mitigate, or isolate known-exploited critical vulnerabilities within 12 hours — a window so tight it renders the traditional monthly patch cycle functionally obsolete. The Computer Emergency Response Team India (CERT-In) released the directive on 26 May 2026 inside a 38-page document titled Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure (reference: CISG-2026-02).
Why 12 Hours?
The numbers behind the directive are stark. According to a Cloud Security Alliance analysis of the blueprint and supporting threat data, the average window between CVE publication and active exploitation in the wild has contracted from roughly 56 days in 2024 to as few as 10 hours by mid-2026 — driven by AI tooling capable of generating working exploits within minutes of a vulnerability's public disclosure. CERT-In states directly that "AI-assisted cyber exploitation reduces the time required for adversaries to identify, weaponize, and exploit vulnerabilities." The 12-hour window applies specifically to known-exploited vulnerabilities on internet-facing and "crown-jewel" systems where feasible. It is not a blanket universal deadline.
The blueprint sets a tiered remediation schedule beyond that headline figure:
Three AI Threat Vectors the Blueprint Names
CERT-In organises the threat landscape around three distinct AI-driven shifts. First, automated reconnaissance and exploit-code generation — AI agents scan entire networks, identify exposed services, and produce working exploits at machine speed. Second, hyper-personalised social engineering: AI-written phishing emails and deepfake voice calls that defeat the human detection heuristics most security awareness training is built on. Third, polymorphic malware that rewrites its own signatures on the fly, evading the detection-by-pattern logic that underpins most endpoint security stacks.
The Register's coverage notes that frontier AI models are actively cited in the blueprint as amplifiers of attacker capability — including examples of criminal groups using AI to build zero-day exploits for planned mass campaigns. The document also addresses prompt injection and model jailbreaking as threat surfaces, reflecting how AI systems themselves have become attack targets, not just attack tools.
What Practitioners Are Saying
Dray Agha, Senior Manager of Security Operations at Huntress, welcomed the explicit carve-out for temporary mitigations: "By explicitly encouraging temporary mitigations, such as isolation, access restriction, or disablement until a patch is ready, this turns the patching deadline into a highly feasible and necessary containment strategy." Agha added that defenders already witness vulnerabilities exploited within hours in the wild — making rapid-response adaptation a practical necessity, not an aspirational target.
What Organisations Must Actually Do
While the blueprint is technically advisory, for India's critical-infrastructure operators it is not optional. CERT-In holds authority to impose the six-hour incident-reporting obligation under the 2022 IT (Amendment) Rules, and non-compliance with that existing rule carries enforcement teeth. The new guidance sits within that regulatory gravity.
The document requires organisations to keep a dedicated inventory of all AI systems that could be either a vulnerability or an attacker's tool. It mandates building what CERT-In calls an "Agentic SOC": a security operations centre that uses AI-assisted tooling for continuous monitoring, alert triage, and rapid response. The three-phase implementation roadmap runs across 60 days: Days 0–7 cover immediate risk reduction (governance, internet-facing asset inventory, activating rapid patching); Days 8–30 focus on operational strengthening (monitoring uplift, AI governance, supply-chain assurance); Days 31–60 target advanced resilience through automation-assisted defence and continuous validation.
Security practitioners have noted publicly that the 12-hour directive does not demand full remediation. Temporary isolation, access restriction, or disabling an exposed service until a patch is ready all satisfy the intent. The point is to eliminate the dwell time that AI-assisted attackers now exploit between vulnerability disclosure and defender action.
Why This Matters Beyond India
India's digital economy is one of the fastest-scaling in the world, and CERT-In's regulatory posture tends to ripple through ASEAN enterprise security teams — particularly in Singapore, Malaysia, and Indonesia, where shared infrastructure, outsourcing relationships, and supply-chain ties create direct exposure to the Indian market's threat surface. A 12-hour patching expectation at India's critical infrastructure layer raises the bar for every vendor and partner connected to it.
The blueprint is also the most detailed AI-threat guidance any major Asian cyber agency has published to date. Previous CERT-In advisories focused on specific CVEs or incident response. This one treats AI-accelerated attacks as a structural shift demanding a change in operations, not just compliance paperwork. Monthly patch cycles built around change-management windows are a design assumption the blueprint explicitly challenges.
The premise held up, and the numbers got worse
Ten hours from disclosure to exploitation was the figure behind the directive. Independent measurement over the following months supports the direction and puts a different shape on it.
CrowdStrike's 2026 threat-hunting work places 88 per cent of observed proof-of-concept exploitation inside two days, with two China-linked groups inside one. That is a distribution rather than an average, and it is the more useful way to plan: a defender is not racing a mean, they are trying to be outside the tail.
The attribution half of the directive also firmed up. A five-agency advisory in August named Siemens S7 programmable logic controllers at water, energy, chemical and manufacturing sites, with exploitation scripts written with AI assistance and disguised as monitoring tools. CERT-In's claim that AI compresses the weaponisation step is no longer an inference from timelines.
The capability is uneven, which is the part worth planning around
Finding a vulnerability and writing a working exploit for it are separate skills, and current models are much better at the first.
Z.ai's GLM-5.3 edges Mythos 5 by 0.7 points on vulnerability detection and trails it by 23.6 on exploit development, on the company's own benchmarks. OpenAI paused its Astra model for two weeks because internal evaluations could not rule out that it had reached the Critical cybersecurity threshold in its own framework, and now runs monitoring on all tool-using inference at roughly 20 per cent of the compute it watches.
For a defender that gap is the current margin. Discovery at scale means more disclosed flaws arriving faster, which is the volume problem a 12-hour window addresses. Weaponisation lagging means the interval between disclosure and a reliable exploit has not yet collapsed to zero. The directive is timed for the moment that gap closes, and the published evidence says it is narrowing rather than closed.
The clock has a starting problem
Twelve hours to patch, mitigate or isolate applies to known-exploited critical vulnerabilities. That phrasing does the work, and it is where the mechanism is exposed.
An unauthenticated SQL injection in GeoServer reaching remote code execution was posted publicly on 12 August. Scanning began within hours, before a patch existed and before the flaw entered any catalogue a vulnerability programme polls. It had no CVE.
A 12-hour clock keyed to a catalogue cannot start until something enters the catalogue. In that case the exposure window opened before the timer could, and no amount of tightening the response interval addresses it. An organisation that follows this directive literally will get very fast at patching catalogued flaws while remaining blind to uncatalogued ones — the reverse of where the risk now lies.
The practical answer is monitoring disclosure channels rather than catalogues, and having a mitigation path that does not depend on a vendor patch existing. Isolate is the third verb in the directive for a reason, and it is the only one available in the first hours of an uncatalogued disclosure.
The supply-chain finding got the least attention and may matter most
The same CrowdStrike work records a DPRK-linked group poisoning 131 trusted AI framework packages.
That attacks the other end of the problem. A 12-hour patch window governs how fast an organisation responds to a flaw in software it knows it runs. A poisoned framework package is a flaw an organisation installs deliberately, through a dependency manager, without any disclosure event to respond to at all.
A patching directive cannot address that kind of supply-chain attack. Any organisation that treats CISG-2026-02 as a complete AI-era security plan is only covering the risks that come with a timer.
View author profile → · Editorial policy
