The Mythos Era, When AI Broke Cybersecurity’s Last Asymmetry

The Mythos Era, When AI Broke Cybersecurity’s Last Asymmetry

For the almost three decades that I have worked in the domain, cybersecurity operated on a reassuring asymmetry: Attackers needed extraordinary skill, time, and patience to find and exploit software vulnerabilities, while defenders, though perpetually behind, could at least count on that human bottleneck to slow the threat. A sophisticated cyberattack required months of reconnaissance, days or weeks of vulnerability research, hours of exploit development, and then a window of opportunity to deploy it before defenders noticed and patched the flaw. The asymmetry was not fairness, but it was predictability. Attackers were human. Defenders could think like humans. Claude Mythos Preview, Anthropic’s frontier AI model announced on April 7, 2026, has shattered that assumption. In controlled evaluations, Mythos Preview could execute multi-stage attacks on susceptible networks, discover and exploit vulnerabilities autonomously, performing tasks that would take human professionals days of work in a matter of hours or even minutes. The reassuring asymmetry is gone. In the age of AI, no system is safe.

The New Capability: Autonomous, Multi-Stage, and Devastating

Mythos Preview has a fundamentally different architecture from its predecessors. It is not merely a faster version of existing AI; it is a different kind of intelligence, optimised for reasoning about complex, multi-step problems. In cybersecurity, this translates into capabilities that were previously the exclusive domain of elite human hackers.

First, Mythos can chain multiple small vulnerabilities into a single devastating attack. A human hacker might find a minor information disclosure flaw in one component, a weak authentication mechanism in another, and a memory corruption bug in a third. Individually, these are low-risk issues. But a skilled hacker can chain them together—using the information disclosure to learn memory addresses, the weak authentication to gain a foothold, and the memory corruption to escalate privileges. Mythos can do this autonomously, scanning for vulnerabilities, evaluating their potential for chaining, and constructing an attack path.

Second, Mythos can reconstruct source code from deployed software. Many software applications are distributed in compiled form, without source code. Finding vulnerabilities in compiled code requires reverse engineering—a slow, painstaking process. Mythos can analyse the compiled binary, reconstruct an approximation of the source code, and then analyse that reconstructed code for vulnerabilities. This is not theoretical; in tests, Mythos found vulnerabilities in widely used software that had escaped human detection for years.

Third, once inside a network, Mythos can automatically map systems, move laterally, and build custom tools to extract data. It does not need pre-written scripts or manual commands. It analyses the network topology, identifies high-value targets, crafts custom exploits for the specific versions of software it encounters, and exfiltrates data—all without human intervention.

The most chilling implication is the democratisation of offensive capability. Engineers with no formal security training can ask Mythos to “find remote code execution vulnerabilities in this web application” and wake the following morning to find a complete, working exploit. What once required years of training, deep knowledge of assembly language, and intimate familiarity with operating system internals is now a prompt away. The exclusive domain of national state actors and elite hacker collectives is now available to anyone with an API key.

Project Glasswing: A Controlled Solution Under Strain

Anthropic’s response to this sobering reality is Project Glasswing, a controlled, invitation-only consortium of roughly 50 organisations, including AWS, Microsoft, Google, Apple, and CrowdStrike, given access to Mythos Preview for defensive security work. Over the past few weeks, Anthropic used Claude Mythos Preview to identify thousands of zero-day vulnerabilities in every major operating system and web browser, along with other important pieces of software. The intent is to get defenders ahead of the curve before Mythos-class capabilities reach less scrupulous hands.

The logic is sound in principle. By giving defensive teams access to the same powerful AI, they can find and patch vulnerabilities before attackers can exploit them. The human bottleneck is replaced by an AI accelerator, but the hope is that defenders, with early access, can stay one step ahead.

However, the execution reveals a structural tension at the heart of modern technology governance. Cybersecurity is no longer centred solely on defending systems against enemies but is increasingly about managing AI systems themselves. The long-standing gap between those who can discover vulnerabilities and those who can exploit them is collapsing. Glasswing addresses the top of this threat pyramid: widely used software maintained by well-funded companies like Microsoft, Google, and Apple. These companies have the resources to participate in Glasswing, to integrate its output into their development processes, and to push patches to billions of users.

But Glasswing leaves the vast underbelly of custom, legacy, and underfunded systems essentially untouched. A small hospital running a custom patient records system written a decade ago is not a Glasswing member. A state government’s tax portal built by a low-bid contractor is not in the consortium. A utility company’s SCADA system running Windows 2000 because it is “too critical to upgrade” is not covered. Project Glasswing, in its current structure, is a form of digital partnership—but a deeply unequal one. This new digital divide is not about access to technology; it is about access to the tools that secure it.

The India Imperative: An Underappreciated Vulnerability

For India, the implications are urgent and underappreciated. The Fintech Association for Consumer Empowerment (FACE) has urged members to reinforce cyber defences and adopt continuous vulnerability solutions and zero-day vulnerability intelligence in response to Mythos’s capabilities. This is a start, but it barely scratches the surface.

India runs enormous volumes of financial, governmental, and civic transactions on software that is old. The core banking systems of many public sector banks date back to the 1990s. The passport issuance system, the railway reservation system, the income tax portal—all run on a mix of modern and legacy code. Legacy code is not necessarily insecure, but it has had less security review, and it is often written in languages that are prone to memory safety vulnerabilities. When Mythos-level tools start finding zero-days in old codebases, Indian institutions will be exposed to vulnerabilities they cannot quickly patch. The average time-to-exploit for a newly discovered vulnerability now sits under 20 hours. Indian IT teams, understaffed and overworked, cannot patch at that speed.

Critically, Project Glasswing does not cover the thousands of custom applications built by Indian banks, government departments, state utilities, and telecom companies. These are not products of Microsoft or Google; they are bespoke systems developed by local vendors, often without rigorous security testing. They are the perfect targets for Mythos-powered attacks.

The article argues that CERT-In (the Indian Computer Emergency Response Team) and the Ministry of Electronics and Information Technology need to urgently pursue AI-assisted security audits of domestic critical infrastructure as a first step. This would involve using Mythos-class AI (or its equivalents from other vendors) to systematically audit the software running India’s critical systems, identify vulnerabilities, and prioritise patches. It would require a national effort, coordinated across multiple ministries and private sector partners. It would require funding, political will, and technical expertise. And it would need to happen quickly.

The Governance Challenge: Integrating Cybersecurity, AI, and Crisis Management

What Mythos and Glasswing together announce is a new epoch. Cybersecurity, AI governance, and crisis management no longer exist as separate disciplines. They must be integrated into one framework of digital risk governance capable of addressing autonomous, probabilistic, and high-impact systems.

A traditional cybersecurity framework assumes that attacks are launched by humans, that vulnerabilities are discovered by humans, and that defenders have time to respond. That framework is obsolete. In the Mythos era, attacks can be launched by AIs, vulnerabilities can be discovered by AIs, and the response time is measured in hours, not days. A probabilistic understanding of risk is needed: not “is this system secure?” but “what is the probability that an AI could compromise this system within a given timeframe?”

Critics have questioned how much of Mythos’s projected danger is clever marketing. Anthropic is simultaneously the creator of the threat and the curator of its solution—a conflict of interest that deserves scrutiny, even if independent evaluations confirm the threat is largely genuine. The company could be exaggerating the capabilities of its model to create demand for its defensive services. It could be downplaying the risks of its model being leaked or stolen. Independent validation is essential. The world cannot rely on a single vendor’s assurances.

Nonetheless, even if Anthropic’s claims are somewhat inflated, the direction of travel is clear. Frontier AI models are becoming more capable at reasoning about complex systems. Sooner or later, a model with Mythos-like capabilities will be widely available—whether from Anthropic, a competitor, or an open-source project. The question is not whether this future will arrive, but whether we will be prepared when it does.

Conclusion: Glasswing Is a Start, But Not Sufficient

Project Glasswing is a possible, necessary first step. It brings together the world’s largest technology companies to use AI defensively. It has already found thousands of vulnerabilities in widely used software. It demonstrates that AI can be used for defence as well as offence.

But it is clearly not a sufficient response. The digital divide in AI security will widen as those with access to Glasswing-class tools pull ahead, while those without—small businesses, developing countries, legacy systems—fall further behind. India, with its vast installed base of legacy software and its strategic importance as a digital economy, cannot afford to be on the wrong side of this divide.

The new epoch demands a new approach. Governments must invest in AI-assisted security audits. International bodies must develop norms for the responsible use of AI in cybersecurity. Open-source security tools must be updated to incorporate AI capabilities. And the private sector must extend access to defensive AI beyond a small consortium of wealthy tech companies.

The reassuring asymmetry is gone. In the age of AI, no system is safe. The only question is how quickly we can adapt.

Q&A: Claude Mythos and the Future of Cybersecurity

Q1: What was the “reassuring asymmetry” in cybersecurity that Mythos has shattered?

A1: For nearly three decades, cybersecurity operated on the assumption that attackers needed “extraordinary skill, time, and patience” to find and exploit vulnerabilities. This human bottleneck slowed the threat. Defenders, though perpetually behind, could at least count on attackers being human—with human limitations, human error, and human time constraints. Mythos Preview has shattered this assumption because it can execute multi-stage attacks autonomously, discovering and exploiting vulnerabilities in hours or minutes rather than days or weeks. What once required elite human hackers can now be done by an AI with a simple prompt. The “reassuring asymmetry” between attacker effort and defender response is gone.

Q2: What specific capabilities make Mythos Preview different from previous AI models?

A2: Mythos Preview has three key novel capabilities:

• Chaining multiple small vulnerabilities into a single devastating attack: It can scan for low-risk issues individually and combine them into a complete attack path, something previously requiring expert human reasoning.

• Reconstructing source code from deployed software: It can analyse compiled binaries, reconstruct approximate source code, and then analyse that code for vulnerabilities—finding flaws that escaped human detection for years.

• Autonomous lateral movement: Once inside a network, it can automatically map systems, move laterally, craft custom exploits for the specific software versions it encounters, and exfiltrate data—all without human intervention.

Additionally, engineers with no formal security training can ask Mythos to find vulnerabilities overnight and receive a working exploit the next morning. The democratisation of offensive capability, once exclusive to national state actors and elite hacker collectives, is now a prompt away.

Q3: What is Project Glasswing, and what are its limitations?

A3: Project Glasswing is Anthropic’s controlled, invitation-only consortium of roughly 50 organisations (including AWS, Microsoft, Google, Apple, and CrowdStrike) given access to Mythos Preview for defensive security work. Anthropic has used Mythos to identify thousands of zero-day vulnerabilities in major operating systems, web browsers, and other software. However, Glasswing has significant limitations:

• It only covers the “top of the threat pyramid”—widely used software maintained by well-funded companies.

• It leaves the “vast underbelly of custom, legacy, and underfunded systems essentially untouched” (e.g., a small hospital’s custom patient records system, a state government’s tax portal, a utility’s SCADA system running Windows 2000).

• It creates a new “digital divide”—not about access to technology, but about access to the tools that secure it.

• It does not cover thousands of custom applications built by Indian banks, government departments, state utilities, and telecom companies.

Q4: Why is India particularly vulnerable to Mythos-class AI threats?

A4: India is vulnerable for several reasons:

• Legacy software: India runs enormous volumes of financial, governmental, and civic transactions on software that is “old.” Core banking systems of public sector banks date back to the 1990s. The passport issuance system, railway reservation system, and income tax portal run on a mix of modern and legacy code.

• Patching speed: The average time-to-exploit for a newly discovered vulnerability now sits “under 20 hours.” Indian IT teams, understaffed and overworked, cannot patch at that speed.

• Custom applications: Thousands of custom applications built for Indian banks, government departments, state utilities, and telecom companies are not covered by Project Glasswing. They were often developed by local vendors without rigorous security testing.

• Lack of AI-assisted audits: India has not yet begun AI-assisted security audits of its critical infrastructure. The article recommends that CERT-In and the Ministry of Electronics and Information Technology aggressively pursue such audits as a first step.

Q5: What governance changes does the article argue are necessary for the new epoch of AI-driven cybersecurity?

A5: The article argues that cybersecurity, AI governance, and crisis management “no longer exist as separate disciplines.” They must be integrated into one framework of “digital risk governance” capable of addressing “autonomous, probabilistic, and high-impact systems.” Specific changes include:

• Governments must invest in AI-assisted security audits of domestic critical infrastructure.

• International bodies must develop norms for the responsible use of AI in cybersecurity.

• Open-source security tools must be updated to incorporate AI capabilities.

• The private sector must extend access to defensive AI beyond a small consortium of wealthy tech companies.

• A shift from deterministic to probabilistic risk assessment: Instead of asking “is this system secure?”, we must ask “what is the probability that an AI could compromise this system within a given timeframe?”
  The article concludes that Project Glasswing is a “necessary first step, but clearly not a sufficient one.” The new digital divide “is not about access to technology but access to the tools that secure it.” India cannot afford to be on the wrong side of this divide.