When Anthropic announced Claude Mythos Preview on 7 April 2026, it did something unprecedented. It built a frontier AI model so capable at discovering and exploiting zero-day vulnerabilities that it chose not to release it to the public.
Instead, through the newly launched Project Glasswing, Anthropic granted limited, invitation-only access to roughly 40 to 50 organisations responsible for critical software infrastructure, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.
The company committed up to 100 million USD in usage credits to help these defenders identify and patch thousands of previously unknown high-severity vulnerabilities across every major operating system, web browser, and widely used application. Some of those flaws had remained undetected by human experts for up to 27 years.
This was not marketing hype. Anthropic’s own testing showed Mythos Preview achieving a 72.4 percent success rate in generating working exploits, a dramatic leap from near-zero performance in earlier models like Claude Opus 4.6. The decision to withhold general release sent a clear message: the offensive potential of such AI now outpaces our collective defensive capabilities, raising risks to global critical infrastructure.
For fintech leaders, CISOs, CIOs, regulators, and policymakers in Southeast Asia and beyond, Mythos is not a distant laboratory curiosity. It is a wake-up call that the traditional cybersecurity paradigm, built on detection, patching, and eradication, has reached its limits in an AI-accelerated, hyper-interconnected world. We must evolve.
The Limits of the Old Playbook
Traditional security approaches assume threats are relatively slow, human-driven, and containable through signatures, perimeters, periodic patches, and zero-trust verification. These methods still work reasonably well against known, legacy threats. But they were never designed for the scale and speed of AI-generated attacks.
In today’s reality, a single Mythos-class model can autonomously audit millions of lines of code, simulate complex exploit chains, and surface zero-days faster than any human red team. When these capabilities proliferate, and they will, a vulnerability in one widely used library or cloud service can cascade across global supply chains, payment networks, trading platforms, and regulatory reporting systems.
The status quo already delivers far more instability than we openly admit. IBM’s 2025 Cost of a Data Breach Report shows the global average time to identify and contain a breach fell to 241 days, the lowest level in nearly a decade. Mandiant’s M-Trends 2025 report indicates that the global median dwell time for intrusions rose slightly to 11 days, with external notifications often taking 26 days or longer. These figures represent normalised, under-discussed costs: prolonged exposure, repeated minor incidents, patch-induced outages, and silent degradations that rarely make headlines but erode trust and resilience in financial systems.
We tolerate this hidden fragility because it fits familiar compliance checkboxes and audit trails. Boards and insurers reward deterministic controls and reasonable security practices. Yet this approach offers zero scalable defence against the unknown unknowns that frontier AI now generates routinely.
Why a Biological Paradigm Is Now Essential
Nature has solved similar problems over billions of years. Biological immune systems do not pursue perfect eradication of every pathogen, an energetically wasteful and often impossible goal. Instead, they tolerate, compartmentalise, incorporate harmless elements as memory, and co-evolve with threats. Approximately 8 percent of the human genome consists of ancient endogenous retroviruses, some of which now play regulatory roles. The microbiome includes countless entities we live with rather than destroy. Defence is layered, distributed, adaptive, and memory-based.
Evolve by design translates this lesson into cybersecurity. We shift from zero-trace eradication to bounded incorporation: treat threats as evolutionary pressure while keeping impact within acceptable probabilistic bounds, what I term martingale error. In simple terms, martingale error means we accept small, probabilistically bounded disruptions (for example, brief, isolated performance dips) rather than demanding impossible zero-degradation perfection. The goal is antifragility: systems that improve from stress rather than merely surviving it.
Two interlocking mechanisms power this shift:
- Chaos Engineering as Immune Stress Testing Popularised by Netflix’s Chaos Monkey and its successors, chaos engineering intentionally injects controlled failures, simulated exploits, and live-like probes into production or shadow environments. Mature implementations have demonstrably improved availability, reduced mean time to recovery, and built confidence in resilience without catastrophic outages. In the new paradigm, defensive AI agents run continuous, scalable game days at machine speed. Impact stays tightly bounded, for example, transient degradation of less than 0.1 percent latency on isolated nodes for seconds, not system-wide collapse.
- Predatory Transformation as Active Adaptation Autonomous agents do not merely detect or block; they hunt, quarantine, analyse attacker behaviour, and actively transform the environment. Code paths rotate, implementations diversify via evolutionary algorithms, and harmless motifs from threats are incorporated as institutional memory, drawing on decades of research into artificial immune systems (AIS), including negative selection for anomaly detection and clonal selection for adaptation. This aligns with emerging automated Moving Target Defence (AMTD) techniques. Gartner predicts that by 2025, 25 percent of cloud applications will leverage AMTD features and concepts as built-in prevention approaches.
The result is a living, self-improving system rather than a static fortress.
Mapping the Risk Spectrum
This approach handles different threat types more realistically than eradication ever could:
- Unexploited (dormant) zero-days become fuel rather than hidden time bombs. Proactive chaos simulations surface them early. Predatory agents diversify implementations around them, turning potential time bombs into low-cost evolutionary fuel without waiting for the next Patch Tuesday.
- Active attacks in progress trigger real-time predator-prey dynamics. The system mutates live, feeds deceptive data, isolates impact locally, and propagates learned immunity network-wide. Dwell time shrinks dramatically because adaptation happens at machine speed.
- Full-on propagating viruses, ransomware, or worms are compartmentalised like mild endemic infections. The threat provides signal for rapid co-evolution. Instead of chasing perfect removal (which often fails in interconnected environments), the ecosystem incorporates lessons and neutralises future variants faster than attackers can iterate.
Evolution is never unbounded. Software cannot grow wings tomorrow. Hard constraints are essential: maximum runtime overhead (typically less than 5 percent), strict backward compatibility, energy and performance budgets, regulatory compliance floors, and human-defined fitness functions tied to business KPIs. Rollback mechanisms, cryptographic signalling limits between agents, and emergency kill switches prevent runaway behaviour or autoimmunity-style over-reactions. Human oversight remains the ultimate policy layer.
Fintech Risks in Southeast Asia: Why This Matters Locally
In Southeast Asia’s fast-growing fintech sector, these challenges are especially acute. Rapid digital banking expansion in Malaysia, Indonesia, and Singapore has created highly interconnected ecosystems reliant on cross-border instant payment schemes such as DuitNow, PromptPay linkages, and regional QR payment networks. A single zero-day in a shared cloud library or open-source component could cascade through payment rails, lending platforms, and regtech compliance systems. Supply-chain attacks targeting fintech vendors already pose outsized risks in a region where many institutions operate hybrid legacy-modern stacks. Traditional patch-and-eradicate methods struggle with the speed of AI-driven threats precisely when trust and uptime are most critical for financial inclusion and economic growth.
Addressing Legitimate Critique from Traditional CISOs and CIOs
Resistance is understandable and should be engaged seriously. Many CISOs will argue that chaos engineering and predatory transformation introduce unacceptable instability, especially in regulated fintech and critical infrastructure environments. Probabilistic, self-mutating systems are harder to audit, certify, and insure. Short executive tenures (often 18 to 30 months) incentivise risk aversion. Legacy maintenance already consumes 60 to 80 percent of many IT budgets; bold experiments threaten headcount, vendor relationships, and career safety. “We followed best practices and were breached by an AI zero-day” is easier to defend than “our evolving system had an unexpected mutation.”
These concerns are valid, but the status quo already delivers systemic instability, just in familiar, under-reported forms. The new paradigm reframes risk rather than eliminating it: from rare but catastrophic whole-body failure (cascading collapse across interconnected financial rails) to more frequent yet treatable localised infections. The latter is far more survivable, learnable, and containable. Localised impact can be isolated, studied, and used to strengthen the broader system.
Moreover, withholding Mythos from the public does not solve proliferation. State-sponsored teams operate without Anthropic’s safety guardrails and are almost certainly developing comparable or superior capabilities. Sophisticated non-state actors and even skilled individuals with access to advancing open-source models are closing the gap rapidly, likely within the next 12 to 24 months. Static defences hand the initiative to whoever deploys offensive AI first. In fintech, where speed, interconnectedness, and trust are paramount, clinging to archaic approaches is not prudence; it is deferred catastrophe.
A Pragmatic Path Forward, With Policy Support
Implementation must be hybrid and phased, not a big-bang replacement. Begin with shadow canaries and non-critical workloads. Expand consortia like Project Glasswing into broader industry pilots across ASEAN fintech hubs. Maintain traditional controls as the innate immune layer while layering on adaptive capabilities. Define success through measurable resilience metrics: faster recovery times, lower aggregate impact under stress, and bounded degradation, not the unattainable zero-incident ideal.
Policy and regulatory frameworks can accelerate responsible adoption. Regulators could introduce safe harbour protections for organisations experimenting within clearly defined evolutionary bounds. Cyber insurers should shift incentives toward resilience outcomes (for example, quantified recovery speed and martingale-style impact tolerances) rather than pure checklist compliance. Governments and industry bodies can fund shared chaos engineering sandboxes and mandate evolutionary resilience reporting for systemically important financial infrastructure. In Southeast Asia, where digital finance growth is rapid and interconnectedness with global systems is deepening, such forward-looking policy could position the region as a leader in next-generation cyber resilience.
The building blocks already exist: decades of artificial immune system research, proven chaos engineering practices, early AMTD deployments, and the defensive potential of frontier AI itself. Mythos did not create the problem; it simply made the obsolescence of the old paradigm impossible to ignore.
Interconnectedness has long been viewed as cybersecurity’s greatest weakness. Evolve by design flips that script: it turns shared threat signals into collective strength. Like every living system that has survived predation across deep time, our digital infrastructure must adapt within realistic boundaries, or risk being out-evolved by those who weaponize AI without restraint.
The choice is no longer whether to change. It is whether we evolve deliberately and boundedly now, or allow the next Mythos-scale event to force a far more painful transition later.
Featured image: Edited by Fintech News Singapore, based on image by freepik via Freepik
