Late last month, the industry learnt that Anthropic is developing Claude Capybara – also referred to internally as Mythos – a powerful new AI model with substantially improved capabilities in vulnerability discovery, exploit development and multi-step attack reasoning. The details emerged through a data leak rather than a formal launch, but the market response was unmistakable: AI has crossed a critical cybersecurity threshold.
Frontier models are accelerating attack life cycles. They will allow attackers to identify and exploit vulnerabilities at a scale, speed and level of sophistication that until recently was the preserve of advanced nation state actors. For security leaders, this is both a warning and a call to action. It crystallises a trend that has been building for some time: the democratisation and industrialisation of cyberattacks.
Mythos is the early signal of two profound shifts in the threat landscape.
The first is the democratisation of advanced attack capabilities. Techniques that once required elite threat actors or well-funded nation state teams will soon be accessible to low-skilled attackers with AI assistance.
The paths are already clear: abuse frontier models directly, as the Chinese state-sponsored group that used Claude Code to infiltrate roughly 30 organisations last year did, or wait for the same capabilities to appear in open-source models where no usage policies or safety layers stand in the way.
This fundamentally lowers the barrier to entry for sophisticated attacks. Organisations that once considered themselves safe because they were not obvious targets of nation state activity are now at risk from newly capable criminal groups armed with AI-powered tools.
Industrialisation of cyberattacks
The second shift is the industrialisation of cyberattacks. With continued advances in agentic AI, threat actors will be able to scan legacy and software-as-a-service technologies at unprecedented frequency and scale. The result will be a near-continuous flow of novel attack methods targeting enterprise systems, networks and employees.
AI allows attackers to move from manual, artisanal operations to repeatable, automated attack pipelines. Attacks are becoming systematic, scalable and reproducible – more like software manufacturing than craft. This is the era of the AI attack factory.
Read: South Africa Inc must wake up to quantum threat
The convergence of these two forces produces a dangerous outcome: more attackers executing more sophisticated attacks, increasing both volume and velocity simultaneously. The time-to-exploit window is collapsing towards zero.
We should all be alarmed by what the Mythos leak revealed, but we should not be surprised. Security researchers have long anticipated that advanced models would eventually demonstrate proficiency in code review, vulnerability discovery and reverse engineering, and that they would integrate with the tools and APIs that enable penetration testing and exploitation.
The gap between writing code and analysing code is narrower than many realise. An AI system capable of generating sophisticated software can be trained or prompted to identify vulnerabilities within it. Combine that with exploit development and the ability to chain multi-step attacks, and you have an entirely new threat surface.
In response to this evolving threat landscape, security leaders should conduct a rigorous reassessment of their foundations. This is not only about implementing new tools. It is about ensuring that existing tools are actually tuned for the threat that is now emerging.
- A good starting point is to assess the efficacy of the first line of defence. Networks, firewalls, web application firewalls, endpoint protection and e-mail security are all critical, but are they tuned for zero-day protection? Default configurations are rarely optimised to defend against previously unknown exploits. If the perimeter and endpoint estate are running standard baselines, the exposure is real.
- Next, evaluate vendor risk. Look hard at the CVE (common vulnerabilities and exposures) history of the security products on which the organisation depends. When AI compresses exploitation timelines to hours, a pattern of frequent critical vulnerabilities in a security vendor’s own products is no longer a manageable operational burden – it is a strategic liability.
- Hunt the blind spots. Legacy servers, unpatched systems, accounts without multifactor authentication, unprotected remote access – the long tail of enterprise infrastructure is where attacks typically land.
- Accelerate patching cycles and consider automated virtual patching and safe remediation. Time-to-patch becomes increasingly critical as campaign timelines move from weeks to minutes.
- Finally, redefine and reinforce network segmentation to protect the crown jewels. Assume breach, limit lateral movement and ensure critical assets are isolated from general network traffic.
What comes next
The step change in AI models’ offensive capabilities has not happened in isolation. It has arrived alongside a sharp increase in open-source software supply-chain attacks, with both signals pointing to the same conclusion: the speed and surface area of attacks are accelerating.
Whether an organisation has adopted AI or not is irrelevant. Threat actors have, and they will continue to push these capabilities further. New models will keep pushing the boundaries of what is possible, for defenders and attackers alike. That is not a surprise – it is the trajectory the industry has been tracking for years. What the recent disclosures make clear is that continuous reassessment is no longer optional.
- The author, Jonathan Zanger, is chief technology officer of Check Point Software Technologies
Get breaking news from TechCentral on WhatsApp. Sign up here.