Mythos: A challenge for the Indian banking system

When Anthropic unveiled Claude Mythos Preview on 7th of April this year, it described the development as a “watershed moment” for cybersecurity. Within no time, that description has moved beyond a marketing claim and become a working hypothesis for regulators across the world. According to Anthropic, Mythos identified critical vulnerabilities, most of them still unpatched, in every major operating system and web browser against which it was tested. These included a flaw in OpenBSD, the primary OS for firewalls and secure server environments across the globe, that had survived more than two decades of human review. The United Kingdom’s AI Security Institute, which was granted early access, reported that the model succeeded at expert-level hacking tasks more than two-thirds of the time. Anthropic then took the unusual step of withholding general release and instead constituted Project Glasswing, a small consortium of defenders that includes JPMorgan Chase, Microsoft, Google, Apple, Amazon Web Services and Nvidia. The implicit signal to banks across the world is unmistakable: the balance between offence and defence has shifted, and that too, abruptly.

For Indian banking, this is neither a distant concern nor a problem confined to developed-market peers. On 23rd of April, the Union Finance Minister convened a meeting of the heads of banks, along with officials from the Reserve Bank of India and MeitY, to assess Mythos-related risks to the country’s financial system, describing the risk as “unprecedented”. CERT-In followed it up by issuing a high-severity advisory that explicitly referred to the model and urged organisations to treat newly disclosed vulnerabilities as exploitable within hours rather than weeks. The National Payments Corporation of India and some of the banks have, according to public reports, sought early access to Mythos to identify vulnerabilities and ‘zero-day’ cyber risk exposure. These responses have been swift and serious, and confirm what the banking ecosystem is beginning to internalise: a model that no Indian institution can directly access may nevertheless reshape the threat landscape in which it currently operates.

To its credit, the RBI has not been caught flat-footed. The Framework for Responsible and Ethical Enablement of AI, namely the FREE-AI Committee report of August 2025, was the most comprehensive engagement by any Indian regulator with the question of Artificial Intelligence in the financial sector. It was organised around 7 “Sutras” and 26 recommendations across 6 strategic pillars. Its survey findings deserve renewed attention in the context of Mythos. It revealed that out of the regulated entities reporting AI use, only 14 percent conducted real-time performance monitoring of their models, only 18 percent maintained audit logs, and a clear majority sought sharper guidance from the central bank on the use of external large language models. The survey findings reflect that robust efforts are required to enhance the preparedness for reliable and effective AI adoption across the financial sector. MuleHunter, the RBI-developed AI tool, now being pressed into wider adoption by the Department of Financial Services, is another quiet but consequential step. RBI’s discussion paper “Exploring Safeguards in Digital Payments to Curb Frauds”, inter-alia, proposing a ‘cooling-off period’ for higher-value digital transactions, ‘kill-switch’ to instantly block digital payments, ‘trusted person’ approval for vulnerable users etc., whatever their frictional cost maybe, also reflects a willingness to prioritise security where the trade-off so warrants.


What the moment now calls for is the next iteration: translating the FREE-AI architecture from a thoughtful framework into an enforceable, time bound operating discipline. In this context, the following steps merit immediate consideration.

First, the recommendations on AI incident reporting and AI inventories should be moved from advisory status to mandatory compliance through a notified circular, with a clear implementation calendar for scheduled commercial banks, NBFCs and payment system operators. A standardised incident reporting format already exists in Annexure VI of the Committee’s report. What remains absent is regulatory enforceability.

Second, the AI policies of banks, as recommended by the FREE-AI report, should require an explicit section on adversarial AI risk and deepfake-driven social engineering. These are areas where existing IT and cybersecurity policies remain demonstrably under-scoped. The proposed AI innovation sandbox should also be expanded to include an explicitly adversarial dimension, partly modelled on Project Glasswing. The RBI’s Innovation Hub, working with IDRBT and CERT-In, is well placed to host controlled red teaming of bank systems against frontier model class capabilities that individual institutions cannot meaningfully simulate in-house.

Third, cooperative banks, smaller NBFCs, and MSMEs remain the system’s weakest link. The Committee’s own data showed AI maturity in these segments to be effectively negligible. A targeted technology and capacity building fund, perhaps administered through a designated nodal authority, would help these entities meet baseline cyber resilience standards. Without such support, their interface with the wider financial system could become a potential contagion vector.

Fourth, India should consider a formal engagement protocol with frontier AI laboratories to enable early notification of capability releases that may have a bearing on financial stability. Such an arrangement would mirror mechanisms that other jurisdictions are attempting to put in place. India’s data localisation rules may complicate the efforts of domestic financial sector entities to test against Mythos on Anthropic’s foreign-based servers; this may require a conscious carve-out to permit supervised defensive access.

Indian banking has successfully weathered every previous paradigm shift - including core banking, mobile internet banking, UPI and the DPDP Act, etc., with the world watching. The Mythos moment is unique in itself in as much as it represents an exogenous capability shock for the financial sector, that is poised to materially reshape it. The regulatory groundwork has already been laid and, in certain respects, is ahead of comparable jurisdictions. What remains is the discipline of execution.



The writer is a Shareholder Director on the Board of Canara Bank and former Vice President of the Income Tax Appellate Tribunal. Views are personal.