An AI model that finds old flaws, and an Indian regulatory response within weeks
On 7 April 2026, Anthropic published research on Claude Mythos Preview, a model it released only to vetted security teams. By Anthropic's own account, the model found a 27-year-old bug in OpenBSD, a 16-year-old flaw in FFmpeg and a 17-year-old unauthenticated remote root flaw in FreeBSD's NFS server (CVE-2026-4747). In a test on Firefox's JavaScript engine, it produced working exploits 181 times, where the previous model managed two. Anthropic also said that more than 99% of the vulnerabilities it had found were not yet patched when it wrote. These are the vendor's own figures, measured under its own test conditions, but they explain why regulators paid attention.
India moved quickly. As MediaNama reported, Finance Minister Nirmala Sitharaman chaired a meeting on 23 April with bank chiefs and senior officials and called the risk "unprecedented". Three days later CERT-In, India's national cyber agency, published a High-severity advisory that names no single vendor but describes exactly this class of capability. On 5 May SEBI went further and named the model.
Advisory CIAD-2026-0020, in plain English
The advisory, titled Defending Against Frontier AI Driven Cyber Risks, warns that frontier AI systems can analyse source code at scale, find zero-day vulnerabilities, accelerate exploit development, automate reconnaissance and plan multi-stage attacks, work that previously needed teams of skilled people. It adds that such tools "lower the barrier to entry" for attackers, and that their dual-use nature cuts both ways. Its recommendations are practical, and none of them needs an AI budget.
- Assume speed. Treat every newly disclosed critical vulnerability in widely deployed software as something that could be exploited "within hours, not weeks".
- Patch in a day. Treat critical patches as urgent and aim to apply them within 24 hours of release, automating where the outage risk is acceptable.
- Shrink what faces the internet. Remove unnecessary internet-facing services, close unused ports, enable DDoS protection, and keep an inventory of everything exposed.
- Trust nothing by default. Multi-factor authentication on internet-facing assets, least-privilege access, micro-segmentation so one breach cannot spread, and a review of older VPN appliances, which attackers have repeatedly used as entry points.
- Know your dependencies. Keep a current inventory of assets and software, patch open-source components, and ask suppliers to show how they are preparing for faster exploit timelines and whether they scan their own code.
- Use AI on defence too. The advisory recommends adopting AI-enabled defensive tools for automated vulnerability detection, attack-surface analysis and threat detection.
CERT-In's own advice is to fix critical flaws within a day and to use AI for defence. The organisations that find their own weaknesses first have the advantage.
A circular that names Claude Mythos
SEBI's circular of 5 May 2026 (reference HO/13/19/12(1)2026-ITD-1_CIMGI/10873/2026) is titled Advisory on Emerging Advanced Artificial Intelligence (AI) Tools for Vulnerability Detection (like Mythos). It is addressed to AIFs, stock exchanges, depositories, clearing corporations, mutual funds, brokers, custodians, credit rating agencies, merchant bankers, portfolio managers, registrars, KYC registration agencies and others. It says tools such as Claude Mythos may enable identification and potential exploitation of existing vulnerabilities "using speed and scale", and may raise concerns about data confidentiality, application integrity and the reliability of outputs.
SEBI set up a task force, cyber-suraksha.ai, to study the risks, share threat intelligence and review the security of third-party application vendors. Its advisory, to be read with SEBI's Cybersecurity and Cyber Resilience Framework, advises regulated entities to:
- apply the latest patches immediately, and consider virtual patching where no patch exists yet;
- run vulnerability assessments, using conventional and suitable AI-based tools where possible, and carry out security audits on a regular or continuous basis;
- work with their third-party vendors so that patches are released and deployed on time.
This is an advisory, not a new law, and it applies to securities-market entities. But a regulator naming a specific AI model in a circular is a signal that other sector regulators and large clients will ask similar questions of their suppliers.
If you ship software, run an internet-facing system or sell to a regulated client, it applies to you
Most Indian businesses are not regulated by SEBI, but most are somebody's vendor. A software company serving a bank, a hospital group serving an insurer or a state department's technology partner will increasingly be asked the question CERT-In tells organisations to ask suppliers: do you scan your own code, and how fast do you fix what you find?
Three groups feel it first. Companies with legacy code, because old systems hold the bugs that fast AI review is best at finding. Teams that rely on open-source dependencies, because a flaw in a component you did not write is still your exposure. And mid-sized organisations without an in-house red team, which is most of India, including most of Jaipur and Rajasthan's businesses, hospitals, colleges and government vendors.
Powerful, but only with a human in the loop
- It is fast. A model can read a large codebase and probe for flaws in hours, where a manual review takes weeks. Anthropic reports that the OpenBSD finding came from a run costing under $50.
- It also produces false leads. A model's output is a lead, not a verdict. Each finding needs to be reproduced and rated by an engineer before it reaches a report.
- The headline numbers are the vendor's. Anthropic's results come from its own tests on open-source software. Your code is different, so treat them as direction, not a guarantee.
- Where the data goes matters. Cloud models process code on the provider's infrastructure, under its terms. Decide up front what may go to a cloud model and what must stay in your environment.
- It does not replace a mandated audit. Where a regulation, tender or client requires a certified audit by a CERT-In empanelled auditor, only such a firm can provide it. AI-assisted review helps you arrive at that audit with fewer surprises.
AI cybersecurity from Jaipur, powered by Claude Mythos 5.2
Data Ingenious Global Limited now has validated access to Claude Mythos 5.2, and we are launching an AI cybersecurity service to put it to defensive use for Indian and international clients. We are a Registered Partner in Anthropic's Claude Partner Network, headquartered in Jaipur, with 25+ years of building and running government and enterprise systems, including work embedded in the Government of Rajasthan's production platforms on RajSevaDwar and IFMS 3.0.
- Source code security review. Mythos-assisted review of applications and services, with engineer-validated findings.
- Legacy and dependency audit. Deep review of old codebases and open-source components, linked to our modernization practice.
- Vulnerability triage and validation. Turning scanner and bug-bounty backlogs into a ranked, reproduced list.
- Authorized red-team support. Support for penetration tests and red-team exercises on systems you own or are authorized to test.
- AI deployment security. Threat modelling for LLM applications and agents: prompt injection, tool permissions, data leakage.
- Remediation and secure rebuild. The same engineering teams patch, refactor or rebuild, then retest.
We are plain about the limits. Every engagement starts with a written scope and authorization, we test only what you own or are authorized to test, and we do not do offensive work against third parties. We do not claim CERT-In empanelment for this service. Claude and Mythos are Anthropic products, and Anthropic does not operate or endorse our service.
Seven questions to ask any firm, including us
Jaipur has a growing group of vulnerability assessment and penetration testing firms, and national players serve the city remotely. Whoever you consider, ask:
- Is everything scoped and authorized in writing? A serious firm refuses to test anything you have not authorized.
- Are they CERT-In empanelled, and do you need that? Check your regulator, tender or client contract. Where it is required, nothing else substitutes.
- Can we see a sample report? Look for reproduction steps, severity reasoning and fix guidance, not a scanner printout.
- Who validates the findings? Ask whether a named human reproduces each one, especially if AI tools are used.
- Where does our code and data go? Ask which tools see it, where it is processed and how long it is retained.
- Can they fix, or only find? A report you cannot act on has little value. Ask who patches, and who retests.
- What have they kept running? Ask for references from systems that cannot go down, not only from assessments.
Eight things to do this month, none of which needs AI
- Inventory. List every internet-facing system, port and API, plus the software and components behind them.
- Remove. Switch off services and ports you do not need, and put DDoS protection on what remains.
- Patch in a day. Agree an owner and a 24-hour target for critical patches on exposed systems, and automate where outages are tolerable.
- Add MFA. Enforce multi-factor authentication on remote access, cloud consoles, admin panels and third-party integrations.
- Replace old VPNs. Review, harden or retire legacy remote-access appliances.
- Watch for speed. Tune monitoring for rapid scanning, unusual credential use and unfamiliar scripts, and review logs more often.
- Ask your vendors. Ask each critical supplier how they patch and whether they scan their own code.
- Scan your own code first. Run an AI-assisted review of your most exposed and oldest applications before someone else does.
Want to know what an AI-assisted review would find in your systems? Tell us what you want assessed and any compliance or data-handling limits. A security engineer replies to agree a written scope, with no obligation.
Prefer email? Write to business@data.in.
Frequently asked
Claude Mythos is a class of Anthropic frontier AI models with advanced cybersecurity capabilities, released under restricted access to vetted security teams. In April 2026 Anthropic published research on a preview of the model reporting that it found and exploited serious vulnerabilities in widely used software, including flaws that had survived decades of human review and automated testing.
It is a High-severity advisory titled Defending Against Frontier AI Driven Cyber Risks, first issued by CERT-In on 26 April 2026. It warns that frontier AI systems can discover vulnerabilities, chain multi-stage attacks and weaponize flaws at a speed and scale that once needed teams of experts, and it tells organisations to treat every newly disclosed critical vulnerability as something that could be exploited within hours, not weeks.
SEBI circular HO/13/19/12(1)2026-ITD-1_CIMGI/10873/2026 is titled Advisory on Emerging Advanced AI Tools for Vulnerability Detection (like Mythos). It names Claude Mythos as an example of tools that can enable identification and potential exploitation of existing vulnerabilities at speed and scale, sets up a task force called cyber-suraksha.ai, and advises regulated entities to patch immediately, use virtual patching where no patch exists, run vulnerability assessments with conventional and suitable AI-based tools, and push their third-party vendors to release timely patches.
The SEBI circular is addressed to securities-market entities such as brokers, exchanges, depositories, mutual funds and registrars. The CERT-In advisory is addressed to organisations generally, and its recommendations on patching, multi-factor authentication, attack-surface reduction and supplier expectations apply to any business with internet-facing software. Check whether a sector regulator or a client contract also sets requirements for you.
CERT-In advises treating critical patches as urgent and aiming to apply them within 24 hours of release, enabling patch automation on internet-facing systems where the outage risk is acceptable, and prioritizing by which flaws are most likely to be exploited and which systems are most exposed, not by severity score alone.
Data Ingenious Global Limited, headquartered in Jaipur, offers AI cybersecurity services powered by Claude Mythos 5.2, covering source code security review, legacy and dependency audits, vulnerability triage, authorized red-team support, AI deployment security and remediation. Jaipur also has other vulnerability assessment and penetration testing firms. Compare any of them using the questions in this article.
No. Data does not claim CERT-In empanelment for this service. Where a regulation, tender or client requires an audit by a CERT-In empanelled auditor, that audit has to come from an empanelled firm. Data's service helps find and fix flaws before such an audit, and every finding is reproduced and validated by a human security engineer.
If a cloud model such as Claude is used, code or data sent to it is processed on the provider's infrastructure under its terms, and Anthropic applies data-retention rules to its cyber programs. Decide this before any work starts. Where code must not leave your environment, the work can use on-premises models such as ZenithAI and manual expert review for those components.
Use the contact form with the AI Cybersecurity option, or write to business@data.in. Tell us what you want assessed and any compliance or data-handling limits. A security engineer replies to agree a written scope and authorization before any testing begins.
- CERT-In Advisory CIAD-2026-0020, 26 April 2026: Defending Against Frontier AI Driven Cyber Risks (primary source)
- MediaNama, 6 May 2026: SEBI circular HO/13/19/12(1)2026-ITD-1_CIMGI/10873/2026 dated 5 May 2026, with a copy of the circular
- MediaNama, 27 April 2026: Claude Mythos puts India on alert, on the Finance Minister's meeting and sector reactions
- Anthropic, 7 April 2026: Claude Mythos Preview research (vendor-reported results)
- Anthropic: Project Glasswing
- Anthropic: Cyber Verification Program
This article is general information, not legal or compliance advice. Regulatory details are summarised from the cited documents as of the publication date, so check the originals and your own regulator's current requirements. Anthropic's results are vendor-reported.

