A frontier AI model has found thousands of critical flaws in the software that runs the world. The race to patch them before attackers exploit them is now measured in hours, not months.
In early April 2026, Anthropic did something unprecedented for an AI company: it announced a powerful new model and then refused to release it. Claude Mythos Preview, the firm’s most capable frontier system to date, had demonstrated cybersecurity abilities so advanced that Anthropic concluded general availability would pose unacceptable risks to “economies, public safety, and national security.” Instead, Anthropic channelled the model exclusively into Project Glasswing—a coalition of major technology firms, financial institutions, and open-source maintainers tasked with patching critical infrastructure before malicious actors could replicate the same capabilities.
This decision signals a profound shift. For decades, the cybersecurity industry operated on a simple asymmetry: defenders had to protect every possible entry point, while attackers needed only one. Now, AI is accelerating both sides—but the offensive advantage is structural. Attackers gain disproportionately from speed, scale, and automation, while defenders remain constrained by patching cycles, legacy systems, and organisational inertia.
The question facing senior executives, investors, and policymakers is no longer whether AI will transform cybersecurity. It already has. The question is whether institutional responses can adapt faster than the threat landscape evolves.
The Vulnerability Deluge: When Discovery Outpaces Remediation
The traditional cybersecurity timeline followed a predictable rhythm. Human researchers discovered a flaw, reported it through coordinated disclosure, and gave developers weeks or months to develop and deploy a patch before public exposure. This equilibrium has collapsed.
In controlled evaluations, Mythos Preview identified thousands of high-severity zero-day vulnerabilities across every major operating system and web browser. Some flaws had persisted undetected for decades—a 27-year-old vulnerability in OpenBSD survived extensive human review and repeated automated testing. A 16-year-old FFmpeg flaw endured five million automated tests before the model surfaced it.
The volume has overwhelmed existing triage infrastructure. The Cloud Security Alliance documented a near-500% spike in vulnerability submissions in a single month, while some major security groups shuttered intake forms entirely. Linux kernel maintainers saw reports climb from two to ten per week—initially dismissed as “AI slop” or hallucinated findings, but rapidly verified as genuine bugs. The curl project, which discontinued its bug bounty program after drowning in low-quality AI-generated reports, recently reported an increasing number of high-quality, AI-supported security submissions.
This inversion creates a remediation bottleneck. As one industry assessment noted, “the constraint in cybersecurity is shifting from detection to remediation.” Organizations can now find vulnerabilities faster than they can fix them, and the stock of known but unpatched flaws is growing—particularly in open-source and legacy environments where maintenance resources are scarce.
From Discovery to Weaponisation: The Collapsing Exploit Window
The true inflection point, however, is not merely volume. It is the compression of time between discovery and weaponisation.
Historically, a significant gap separated finding a vulnerability from engineering a functional exploit. This window gave defenders critical breathing room.
Advanced AI has collapsed it from weeks to hours.
Anthropic’s technical benchmarks illustrate the shift with stark clarity. When testing against Firefox 147, the previous generation model (Claude Opus 4.6) produced working exploits in just two attempts out of several hundred. Mythos Preview generated 181 working exploits under identical conditions, with 29 achieving full register control. On internal benchmarks involving fully patched targets, prior models achieved low-severity crashes. Mythos reached full control-flow hijack on ten targets.
Industry data confirms the broader trend. The Zero Day Clock, introduced in March 2026, tracks the time between CVE disclosure and confirmed exploitation. The mean time-to-exploit has fallen from 2.3 years in 2018 to approximately 23 hours in 2026.
Real-world incidents already demonstrate what this velocity enables. Sysdig documented an AI-based attack achieving administrator-level access in eight minutes. Security researchers have observed “vibe hacking”—autonomous attack chains where AI scans architecture, identifies zero-days, codes custom malware, bypasses firewalls, exfiltrates data, and drafts ransom notes without human intervention.
The implications extend beyond criminal operations. Nation-state actors and advanced persistent threats now operate with capabilities that were previously resource-intensive and technically demanding. The democratization of offensive cyber power represents a strategic shift with geopolitical consequences.
Critical Infrastructure at Risk: The OpenSSL Precedent
The most consequential vulnerabilities reside not in applications but in foundational libraries—the cryptographic and networking substrates upon which global infrastructure depends.
In January 2026, security researchers at AISLE used AI-driven analysis to discover twelve zero-day vulnerabilities in OpenSSL, the cryptographic library underpinning encryption for the vast majority of internet traffic. Three of the flaws dated to 1998—predating OpenSSL itself, inherited from the original SSLeay implementation—having survived more than 25 years of intensive human and machine auditing.
One vulnerability, CVE-2025-15467, received a CVSS score of 9.8 (Critical)—an exceptionally rare rating for such a mature project. It enabled pre-authentication remote code execution through a stack buffer overflow in CMS message parsing, meaning attackers could hijack execution before any cryptographic verification occurred. No valid keys required.
The April 2026 OpenSSL advisory marked another milestone: for the first time, multiple AI research teams independently converged on the same zero-day. AISLE and Anthropic both identified CVE-2026-28386, an out-of-bounds read in AES-CFB-128, with AISLE reporting it 63 days earlier and developing the patch that OpenSSL adopted. This co-discovery validates the capability while underscoring the competitive dynamics now unfolding—multiple AI systems hunting the same critical codebase simultaneously.
If hostile actors weaponize structural vulnerabilities at this level before patches deploy, the cascading effects could span global banking networks, electrical grids, cloud hyperscalers, and aviation routing systems. The interconnected nature of modern infrastructure means that single-point failures propagate rapidly across sectors and borders.
Project Glasswing: A Defensive Coalition for the AI Era
Faced with these capabilities, Anthropic opted for restriction over proliferation. Project Glasswing represents one of the largest multi-party vulnerability coordination efforts in history, bringing together AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks as launch partners, with access extended to over 40 additional organizations maintaining critical software infrastructure.
The initiative commits $100 million in model usage credits and $4 million in direct donations to open-source security organizations. Partners apply Mythos Preview exclusively to defensive security work—identifying vulnerabilities in their own codebases, developing patches, and sharing findings through coordinated disclosure.
Early results suggest practical impact. Mozilla reported that Mythos identified multiple high-severity vulnerabilities in Firefox. Microsoft noted substantial improvements on CTI-REALM, its open-source security benchmark, compared to previous models. AWS has applied the model to critical internal codebases, while CrowdStrike integrates its trillion-events-per-day sensor network with frontier AI capabilities.
However, the initiative faces inherent limitations. The world’s exploitable attack surface vastly exceeds what any curated partner ecosystem can cover. Most organizations maintaining critical software lack early access to Mythos-class capabilities. Meanwhile, comparable capabilities are proliferating rapidly—OpenAI has expanded its Trusted Access program, and open-weight models are democratising vulnerability research at accessible cost.
The defensive advantage conferred by early access is therefore time-limited. As the Cloud Security Alliance observed, “the capabilities seen in Mythos will quickly become more widely available,” dramatically increasing the frequency and sophistication of attacks organizations will face.
The Nordic and European Dimension: Regulation Meets Reality
For Nordic and European business leaders, the AI vulnerability crisis intersects with an evolving regulatory landscape. The EU AI Act, entering full force in August 2026, introduces automated audit, incident reporting, and cybersecurity requirements for high-risk AI systems.
This creates a liability inflection point. Existing regulations employ “reasonableness” as a standard of care. When AI can discover significantly more vulnerabilities at accessible cost, the definition of reasonable defensive effort shifts. Boards will face questions about whether they deployed available AI tools for defensive scanning, and whether failure to do so constitutes negligence.
Nordic enterprises, with their high digitalization rates and reliance on cloud infrastructure, face particular exposure. The region’s advanced banking networks, energy grids, and telecommunications systems depend on the same open-source cryptographic and networking libraries now under AI-assisted scrutiny. The concentration of critical infrastructure in digitally mature economies creates concentrated risk alongside concentrated opportunity.
Moreover, the Nordic model of collaborative governance—spanning public-private partnerships and cross-border coordination—offers structural advantages for collective defence. The Cloud Security Alliance explicitly recommends that defenders “engage now with sector coordinating groups, ISACs, CERTs, and standards bodies to share threat intelligence, coordinate response, and produce sector-specific guidance.” Nordic institutional frameworks are well-positioned to operationalize this recommendation.
Strategic Implications: What Leaders Must Do Now
The transition to AI-accelerated cybersecurity demands action across three-time horizons: immediate operational response, medium-term program adaptation, and long-term strategic repositioning.
Immediate: Prepare for the Patching Surge
Organisations should anticipate a flood of critical patches from Glasswing partners and open-source projects. Security teams need triage and deployment capacity scaled for continuous, high-volume remediation. The Cloud Security Alliance recommends running tabletop exercises for multiple simultaneous high-severity incidents within the same week, with pre-authorised containment playbooks.
Medium-Term: Adopt Defensive AI at Scale
The same agentic AI capabilities accelerating threats can augment defence. Leaders should formalise AI agent usage across security functions—vulnerability scanning, code review, incident response, and governance—while implementing robust oversight controls. The barrier to adoption is lower than commonly assumed; modern coding agents are now more accessible than traditional enterprise software, and their utility extends far beyond code to GRC, audit automation, and threat analysis.
Long-Term: Architect for Resilience
When vulnerability discovery outpaces remediation, containment becomes the primary control. Organizations must prioritise network segmentation, zero-trust architectures, egress filtering, and phishing-resistant multifactor authentication. These fundamentals limit blast radius when—not if—individual vulnerabilities are exploited.
The Cloud Security Alliance frames the challenge directly: “We cannot outwork machine-speed threats. Re-prioritise, automate, and prepare for burnout.” Security team resilience, including sustainable workload management and retention of scarce expertise, should be treated as a strategic priority with the same urgency as technical controls.
Conclusion: The New Calculus of Digital Risk
The AI vulnerability crisis represents a structural shift, not a temporary disruption. The cost and capability floor for offensive cyber operations is dropping. The time between disclosure and weaponisation is compressing toward zero. Capabilities that previously required nation-state resources are becoming broadly accessible.
For business leaders, this recalibrates the risk equation. Cybersecurity is no longer a cost centre managing probabilistic threats. It is a strategic function operating in a permanently accelerated environment where the baseline assumption must be that unknown vulnerabilities exist in critical dependencies, that they will be discovered by AI systems, and that exploitation may precede patching.
The organisations that thrive will be those that treat AI not merely as a threat vector but as a defensive imperative—deploying machine-speed capabilities against machine-speed threats, hardening architectures for inevitable breaches, and building institutional resilience that transcends any single vulnerability or patch cycle.
Project Glasswing is a necessary first response, but it is not sufficient. The ultimate measure of success will be whether the global technology ecosystem can evolve patching, coordination, and governance mechanisms that match the pace of AI-driven discovery. The alternative is a world where the internet’s foundational software remains permanently vulnerable—and where the next crisis is not months away, but hours.
The Nordic Business Journal provides analysis at the intersection of technology, policy, and strategy for senior executives and investors across Northern Europe and global markets.