Anthropic’s release of the frontier AI model Claude Mythos Preview has caused significant trepidation, as its autonomous code-reasoning capabilities surfaced previously undetected vulnerabilities across major operating systems and software infrastructure. The architecture of access, being constructed around such capabilities, creates an asymmetry in access. India needs to secure critical infrastructure by updating, modernising or replacing legacy systems.

In April 2026, Anthropic stated that its frontier artificial intelligence model, dubbed Claude Mythos Preview, had unprecedented cyber capabilities that made it too dangerous to release publicly.[1] They disclosed that the Large Language Model (LLM), during its testing phase, had identified ‘thousands of high and critical severity vulnerabilities’, many of which have been, and continue to be, validated by cybersecurity experts. These include discovering and exploiting a 27-year-old Zero Day Vulnerability (0-day) in OpenBSD, finding a 16-year-old bug in media processing library FFMpeg, and patching 271 vulnerabilities in Firefox 150 based on a single evaluation cycle.

While Mythos Preview’s capabilities mark a significant step in the development of AI systems, their potential has rung alarm bells among global cybersecurity communities, financial institutions and governments. The global financial elites have especially highlighted cyber threats posed by Mythos. Anthropic has disclosed that Mythos was used to conduct what could essentially be called ‘digital robbery’ on systems across the world, which has led to fears that the system may further be misused for mass looting of bank accounts,[2] paralysis in international payment systems, or spark a crisis of confidence in extant financial systems. [3]

Global finance leaders and central bankers from countries such as the US, Canada, the EU, the UK, Germany, South Korea and India have expressed concerns and interest in accessing the LLM to plug vulnerabilities in their financial systems. Simultaneously, Anthropic has launched Project Glasswing, an AI consortium comprising companies such as Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, as well as over 40 other organisations with access to the system.[4]

The discussions around Mythos and its capabilities are being presented as a singular new threat, a framing that does hold some merit, but does not present a complete picture. Mythos did not create the vulnerabilities it found and developed exploits for; zero-day vulnerabilities have always existed and have been routinely exploited. What has changed in Mythos is the tempo of these threats, as it accelerates the weaponisation of extant fragilities. It is therefore crucial to examine how the capability jump, cost dynamics, human oversight, access distribution, and structural issues interact with the cyber-capabilities Mythos has heralded.

Contextualising the Global Furore

The case of OpenBSD has effectively illustrated the reason for the alarm at Mythos. The flaw had remained undetected in one of the ‘most security-hardened operating systems ever built’[5] since around 1998, despite hundreds of code revisions, security reviews and audits. The delay in discovery may be due to multiple factors: it required a very specific sequence of reasoning among thousands of possible lines of reasoning and permutations. A human auditor may miss it, not due to incompetence, but because the sheer volume of code and the complexity of interactions between components may tax the limits of human cognition. Additionally, programmers often triage the code based on perceived criticality and personal experience, thereby prioritising newer code over stable, mature codebases. Mythos has no compunctions about dealing with legacy code.[6]

Mythos have added considerable dimension to the potential of AI in cybersecurity. While current AI systems do have a degree of autonomy, they require human guidance to reach the desired result. Mythos found over 271 vulnerabilities in Firefox in April 2026 without additional guidance after the initial command.[7] By comparison, Anthropic’s Claude Opus 4.6 found only 22 bugs in a similar study, all of which required human steering. The more consequential capability shift, however, is how Mythos’s autonomy applies to what it finds. Unlike current frontier models, Mythos not only identifies vulnerabilities but can also independently act to close the loop from code analysis to the development of working exploit chains, with minimal human involvement.

Typically, finding a bug and developing an exploit are distinct stages, and the time and capability required to bridge the gap give defenders a window to patch before weaponisation. Mythos, however, collapses this window to the same timescale as discovery itself.[8] The dual-use nature of AI technologies, such as Mythos, means that the same capability that enables a defender to audit a codebase can also enable an attacker to weaponise its findings. Furthermore, reports suggest that during evaluations, Mythos appeared to know it was being tested, showing signs of this awareness in roughly 29 per cent of transcripts, even when it didn’t explicitly say so. It may also have intentionally underperformed to avoid flagging behaviour that might seem disingenuous, thereby indicating a nascent capability to manage perception by manipulating responses.

Cost is the second vector to consider. The scan that found the OpenBSD vulnerability used less than US$ 50 in resources and was part of a broader 1,000-scaffold run that cost US$ 20,000.[9] Bounties for comparable bugs typically run much higher, requiring weeks or months of dedicated work. Cybersecurity experts also tend to be paid significantly higher salaries, based on their expertise and market dynamics. This has two major implications: one, AI systems like Mythos may be leveraged to find critical vulnerabilities at a fraction of human resources. Two, leveraging AI for vulnerability discovery may introduce predictability in cybersecurity costs.[10]

Broadly, the two factors above have significant implications for the human role in AI-enabled processes, particularly in cybersecurity. The potential for lower, more predictable costs due to access to AI capabilities (at least as good as, or comparable to, what Mythos claims to have), especially with higher-order autonomy, has implications for how much human capabilities can match the resulting collapse in timelines. As of now, there is still a need for human expertise in identifying and verifying vulnerabilities; Anthropic itself has launched a public bug bounty programme on the threat-exposure management company HackerOne, inviting external human researchers to identify vulnerabilities in its systems.[11]

However, there is a need to be aware of the possibility that rising autonomous AI capability to find a vulnerability and execute the result (patching and/or developing an exploit) autonomously, combined with LLM systems’ emerging capability to predict and manipulate their responses, may not only weaken any meaningful assertion of human oversight but also significantly reduce the need and demand for human expertise.

The Differential Architecture of Access

Anthropic’s response to Mythos’s capabilities has been to create an AI........

© IDSA