The Mythos menace | Claude Mythos

In an old fable from the Arabian Nights, a poor fisherman frees a genie from captivity only to discover the creature may turn on its liberator. Dario Amodei, chief executive of Anthropic, appears determined not to risk a similar fate. Or so goes the modern fable. His company chose not to publicly release ‘Mythos’, the newest and most powerful model in the Claude family of artificial intelligence (AI) systems, because it was considered too potent for open deployment. Anthropic itself described the model’s capabilities as “substantially beyond those of any model we have previously trained”.

That alone would have made headlines. But what set off alarm bells globally was what Mythos was built to do. Brewed in the cauldron of agentic AI, especially around coding tasks, it’s said to have hyper-evolved into a Frankensteinian monster of sorts, with a transcendent eye for software vulnerabilities, including unknown “zero-day” flaws or bugs unknown to its developer but exploitable by attackers. In short, a capability that could be weaponised, chaining multiple weaknesses into a single crippling attack and potentially destabilising systems at machine speed. Mythos, Anthropic says, has identified vulnerabilities across major operating systems and browsers, including one that had reportedly gone undetected for 27 years. Reports of a possible breach of Claude Mythos Preview by a private forum only deepened the anxiety.

Anthropic’s answer to the risks posed by its own creation was Project Glasswing, a restricted-access cybersecurity initiative that underscored the gravity of the threat. Instead of releasing Mythos openly, Anthropic placed it in the hands of a select consortium of software giants and critical institutions to use it defensively—to find zero-day vulnerabilities, stress-test systems and patch weaknesses before adversaries could exploit them. If Mythos symbolised the danger of autonomous cyber AI, Glasswing was the first attempt to keep the genie in the bottle.

THE WORLDWIDE WORRY

For the rest, we have a technology story turned into a systemic risk one. Across capitals, concern spread that Mythos represents the beginning of a new category of digital threat: not AI that merely generates content or writes code, but AI capable of autonomous cyber operations. The World Economic Forum’s Global Cybersecurity Outlook 2026 captured that anxiety when it warned that while generative AI had so far mainly amplified social engineering and reconnaissance, the emergence of autonomous AI agents capable of executing attacks could mark a turning point. The warning moved rapidly into policymaking. On April 7, US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell reportedly convened an urgent closed-door meeting with chief executives of major American banks to discuss the cybersecurity risks posed by Mythos. Across Europe and Asia, central bankers, regulators and security agencies began examining whether cyber vulnerabilities could evolve into systemic financial risks.

India responded with unusual urgency. On April 23, Union finance minister Nirmala Sitharaman held a high-level meeting with heads of commercial banks and representatives from the Reserve Bank of India and the Indian Computer Emergency Response Team (CERT-In). The finance ministry stressed the need for a robust, real-time threat intelligence-sharing mechanism. Present in the meeting was also Ashwini Vaishnaw, Union minister for electronics and information technology (MeitY), who too urged Indian companies working on AI to build a software model that could counter Mythos’s threat. At a recent media conference, Sitharaman said, “The cyber challenge we have because of Mythos is a big one. MeitY is seized of it and is engaging with the US administration, Anthropic and vendors who have been given a chance to test Mythos to see what steps India should take.” CERT-In warned that frontier AI systems with advanced cyber capabilities could lower the barrier for malicious actors, automate exploitation workflows and scale cyber campaigns.

The urgency of the situation has triggered alarm through many layers of officialdom. “It may affect markets, banking and payments,” a senior government official said. “The concern is not just breach, it is system-wide impact.” That is not language used lightly. Officials have reportedly sought access to better understand Mythos’s capabilities. Regulators have been asked to engage counterparts in the US, UK and Europe. The Prime Minister’s Office is said to be tracking developments closely, with CERT-In and the National Security Adviser’s office coordinating security guardrails. “With Mythos knocking on the doors, the need for deeper coordination across global bodies has become urgent,” an official said.

WHY THE PANIC

Earlier AI tools could assist humans. They could generate code, automate tasks, help defenders or even aid attackers. Mythos is seen as going further by combining reasoning, vulnerability discovery and autonomous execution. Nasscom chairman Srikanth Velamakanni, who is also the co-founder and Group CEO of Fractal Analytics, a frontline AI company, says, “Earlier, even though they were good at reasoning and answering questions, they were not really able to accomplish any real-world tasks with great accuracy. Now, that’s changing. Not only can they plan tasks like suggesting airline flights but can now also access networks and buy the tickets for you. Which means they have more agency—the ability to autonomously decide, access the real world and get things done.”

Much of the concern around Mythos, therefore, comes down to one word: agency. The core fear is that autonomous AI may be able to chain multiple small vulnerabilities into a single crippling attack. Once inside a network, such systems could map architecture, move laterally and generate tools to extract data at extraordinary speed.

Between 2025 and early 2026, frontier models have become significantly more agentic. “These are not meant as cyber-attack systems. They are meant to help solve real-world problems,” Velamakanni says. But that same ability to find ways into systems gives them a dual-use edge. That is what worries governments. “Hidden flaws or zero-day vulnerabilities are marketable commodities. An AI tool can find these vulnerabilities in systems,” he says. And it can do so far faster than before. Experts say that what once took months may now take days or even hours. That compression of time may be the single biggest disruption. Because cybersecurity has always depended partly on defenders having time to beat hackers at their game.

What happens when that time disappears? Rishi Verma, head of the AI Centre of Excellence of Financial Software and Systems (FSS), says Mythos differs because it operates through a “planner and executor loop”, wherein the system not only decomposes goals into tasks that it plans but also executes them iteratively with feedback-driven refinement. “In this sense, if guardrails are insufficient, it could autonomously execute unintended actions at scale.” That makes risk non-linear and cascading and also makes it extremely difficult to predict how far an attack can spread, which system it may impact simultaneously, and how it will propagate across various digital networks.

A vulnerability in municipal software could theoretically be linked to power, water or transport systems. What worries governments is not one exploit but the possibility of multiple systems being probed and disrupted together. Munjal Kamdar, partner, cyber-strategy of Deloitte, frames the issue through “discovery and exploitation”. Discovery is identifying vulnerabilities; exploitation is weaponising them. Historically, there was often a lag between the two. Now, that lag is collapsing. “With models like Mythos, this timeline from a few days could shrink to just a few hours,” he says. Cybersecurity, in effect, becomes a race against time.

INDIA’S VULNERABILITIES

If there is one country that embodies the promise and the risk of Mythos, it is India. Its financial system is vast and deeply software-driven. Banking deposits stand at Rs 251.9 lakh crore; mutual fund assets at Rs 73.7 lakh crore. UPI now has 500 million active users processing 22.6 billion transactions monthly. Few countries have digital financial infrastructure at this scale. But scale creates exposure. “Banking is a highly regulated sector and is sitting on legacy infrastructure,” says Kunal Nandwani, co-founder of uTrade Solutions. Some core systems still run on programming languages such as COBOL, dating back decades. Legacy infrastructure has long posed risks. “AI changes the speed, scale and sophistication with which those risks may be exploited,” he points out.

Compounding this is the structure of modern finance itself. Banks no longer operate in silos. They sit within ecosystems of payment providers, fintechs, insurers, application programming interfaces (APIs) and third-party vendors. “More API integration increases the opportunity for cybercriminals to launch attacks because while banks are regulated entities that follow strict security principles, other financial institutions may not have them,” says Verma of FSS. That is why experts see fintech firms and payment aggregators as particularly exposed. They may be used as entry points into larger banking networks. NBFCs, with heavy third-party dependence, face similar concerns. Neobanks add another layer of distributed risk.

Some industry voices argue that India’s permissioned banking networks, where access is restricted to verified participants, and layered firewalls provide resilience. They may well do. But there are many who warn of the dangers. “Mythos has assembled a series of parallel agents and each of these agents can co-relate with the chain of security events and launch an agent within minutes,” says Ramesh Lakshminarayanan, group head, IT, and chief information officer at HDFC Bank. “Therefore, all vulnerabilities on legacy applications—including zero-day vulnerabilities—can be picked up in a matter of minutes.“

Increasingly, the risk is not viewed as a breach problem, it is viewed as one of contagion. That is where India’s digital public infrastructure enters the picture. UPI, Aadhaar and DigiLocker have transformed the economy. But their interdependence also means vulnerabilities in one area may ripple into others.

There is another anxiety. Some experts privately note that frontier defensive systems such as Glasswing-style capabilities may remain concentrated among top-tier global software players, while legacy-heavy institutions elsewhere face the threat without comparable access to cutting-edge defensive tools. That asymmetry matters. It raised a strategic concern among policymakers: if such frontier defensive tools remain concentrated among a handful of advanced institutions, countries like India may have little choice but to rapidly build sovereign cyber defences of their own.

Unlike traditional cyber threats dependent on human expertise and long timelines, AI-driven vulnerability discovery could drastically reduce response windows for critical systems, from communications to radar. Consumer industries face risks from intellectual property theft. And capital markets may be especially vulnerable. India’s recent clash with Jane Street Capital has shown how sophisticated algorithm-driven strategies can distort markets even without autonomous AI. The New York trading firm was temporarily barred from Indian markets over allegations of a “sinister scheme” to manipulate the Bank Nifty index, with regulators claiming it made thousands of crores in unlawful gains by moving index levels intraday to profit from derivatives. Jane Street has denied any wrongdoing. Imagine such strategies becoming self-learning and operating across multiple institutions simultaneously. That is what has spooked money and capital markets globally. Regulators increasingly view this as digital systemic risk, one that does not replace traditional risks but amplifies them unpredictably.

NEW THREATS, NEW DEFENCES

Finance may be the immediate concern, but Kamdar points to public infrastructure, identity systems, utilities and healthcare as exposed sectors. Hospital networks increasingly integrate patient records, billing systems and medical devices. Pharmaceutical companies hold enormous stores of sensitive data. Compromise here is not merely commercial risk, it can become public risk. The concerns are even sharper in the defence sector (see Code Red).

Experts broadly agree that old cyber doctrines are inadequate. “You cannot bring a sword to a gunfight,” says Nandwani. “We need to defend using smart AI tools.” That phrase captures the emerging consensus. The answer to AI-driven cyber threats may have to be AI-driven defence. Globally, regulators at the G20, Bank for International Settlements and Financial Stability Board have begun discussing frontier AI as a potential source of instability—through cyber contagion, algorithmic volatility or risk transmission. India has been part of those discussions. What has changed, officials say, is the urgency. Within the finance ministry and the RBI, work is under way to map vulnerabilities ranging from financial exploitation to liquidity stress and cross-border contagion. “Transmission across markets could be rapid,” an official said.

That is pushing regulators beyond broad AI principles toward concrete guardrails: tighter oversight of AI-driven trading, real-time audit trails, broker-level safeguards, adversarial red-teaming and AI-led surveillance. Taken together, it marks a shift from reactive oversight to system-wide risk management. Markets may increasingly need monitoring—and, if required, stabilisation—at machine speed. That is a profound doctrinal change.

Yet experts stress Mythos is not only a warning, but also a call for digital defence. During geopolitical incidents such as Operation Sindoor, Kamdar notes how some firms faced 1,000-2,000 attacks per second, making manual response impossible. Automation became essential. “AI must be used to manage and fight the risks posed by AI.” That principle now runs through discussions around “DevSecOps” (development, security and operations), i.e. integrating security testing at every stage of the software development process, and “findings lifecycle management”—the speed at which vulnerabilities can be identified and fixed before exploitation. “Mythos brings a new set of vulnerabilities and unprecedented challenges, and the industry can no longer afford to be complacent,” says Sugandh Saxena, CEO, Fintech Association for Consumer Empowerment (FACE), a self-regulatory body for the fintech sector.

THE BIGGER CHALLENGE

For banks and insurers, says Sony Anthony, partner and head, cyber defence, at KPMG India, the collapse in response windows changes everything. “Previously, it could take up to 2.5 years for a vulnerability to be exploited; now, with Mythos, this timeframe has collapsed to about 20 hours.” To address these risks, Anthony says organisations should consider implementing faster patch cycles, bigger budgets and deeper use of automation.

Vinayak Godse, CEO of the Data Security Council of India (DSCI), argues that the response must go beyond patching. “The world for which we built current security systems no longer exists.” His prescription includes aggressively testing internet-facing applications, reducing attack surfaces and redesigning architecture through techniques such as micro-segmentation. “Until now, finding a zero day was elite work,” he says. “That barrier is now democratised.”

That may be the most unsettling line in this debate. CERT-In has urged firms to elevate alert levels, reduce exposed attack surfaces and adopt AI-enabled defensive tools. It has even advised monitoring outbound traffic to AI services and training security teams to detect how AI-augmented attackers operate. FACE has urged members to accelerate patching, improve zero-day intelligence and ringfence critical systems. The Indian Banks’ Association is similarly focused on hardening payment rails. But experts say India also needs something broader: sovereign cyber capability, deeper engagement with frontier software providers and a long-term national cyber defence strategy. That means not merely reacting to frontier AI but building capability to shape the defensive frontier itself.

HDFC Bank’s Lakshminarayanan says that banks need to start looking at security engineering models where AI reviews software for security vulnerability right when it is developed. Banks can also adopt micro-segmentation to isolate key functions, preventing simultaneous compromise, and use white-box cryptography (WBC), which combines encryption and obfuscation methods to embed secret keys in application code. “A fourth method is hardware-based key routing, where the keys are not in memory but secured in hardware,” he adds.

Ultimately, the Mythos debate is about more than one AI model. It is about whether cybersecurity, finance and governance can keep pace with frontier technologies. Calls for self-regulation have struggled. In 2023, leading AI researchers and CEOs called for pausing more powerful AI systems. Technology moved on. That is why experts increasingly argue global frameworks may be unavoidable, whether through the United Nations, financial regulators or AI safety coalitions. “AI regulation should happen on a global scale,” says Nandwani. But regulation alone will not be enough.

India has built one of the world’s most ambitious digital financial ecosystems. Protecting it now requires treating cybersecurity not as compliance but as economic security itself. The Centre has moved early in recognising the threat. But safeguarding the trust millions have placed in these systems will require much more than advisories. It will require new doctrines, new defences and a new urgency. Because in the age of autonomous cyberAI, the next systemic shock may not begin with bad loans. It may begin with bad code.