For decades software security relied on two assumptions: finding and exploiting vulnerabilities required rare expertise, and defenders had time to assess and fix what was found. With frontier AI those assumptions no longer hold. Flaws can now be discovered at a scale and pace that most security, development, or infrastructure teams were not built to absorb, and attackers frequently exploit them faster than organizations can patch.
This analysis is a collaborative effort by Charlie Lewis, Florian Niedermann, Giulio Romanelli, Martin Harrysson, and Sven Blumberg, with Larry Villella and Wolfram Salmanian, representing McKinsey’s Risk & Resilience Practice and Tech & AI group. It explains why a new approach to software security is necessary, the four levers that must move together, and how organizations can start implementing them.
Evidence of the shift
Frontier models can now read software configurations and code with a security mindset, produce working exploits, or deploy agent fleets to find and exploit vulnerabilities. After Anthropic launched Project Glasswing in April 2026, it reported in May that partners had found more than 10,000 high‑ or critical‑severity vulnerabilities using Claude Mythos Preview; several partners said their bug‑finding rates increased more than tenfold. Among the discoveries were a 20‑year‑old Firefox bug and a 27‑year‑old OpenBSD flaw.
The public software ecosystem shows similar changes: according to a kernel maintainer, the number of CVEs fixed per Linux kernel release rose from roughly 500 across versions 6.9–6.19 to more than 1,000 from version 7.0. In April 2026 HackerOne reported a 76 percent year‑over‑year rise in vulnerability submissions while the share of valid findings remained steady.
In October 2026 two critical PHP vulnerabilities (PHP2Shell) and a high‑severity Node.js flaw (Node2Shell) that could enable remote code execution were identified; PHP2Shell exploitation attempts have been observed. Earlier, CrowdStrike reported that 2025 saw an 89 percent year‑over‑year increase in attacks by AI‑enabled adversaries.
Speed and economics favor attackers
Attackers are moving faster: the average time between disclosure of a critical vulnerability and active exploitation is now hours, whereas a year ago it was measured in weeks. By 2026, about two‑thirds of exploited vulnerabilities were attacked on or before the day of disclosure. Defenders, by contrast, have slowed: Verizon’s 2026 Data Breach Investigations Report found the median time to full remediation rose from 32 to 43 days, with only 26 percent of issues fully remediated.
Practically, this gap forces teams to choose blunt mitigations—often turning off exposed services and accepting operational impact—because processing large numbers of alerts and fixes by traditional, manual means no longer scales. Leading organizations are closing the speed gap by automating detection, prioritization, and containment where possible, while keeping humans in the loop for critical decisions.
From a cost perspective, attackers can ask models to search for vulnerabilities cheaply and quickly. Defenders face higher costs: comprehensive scans with frontier models and the consequent remediation can be slow and expensive. Early adopters report first scans of crown‑jewel systems finishing in days to weeks, with remediation taking months.
The financial cost is only one part of the burden. Fixes compete for scarce engineering capacity and change windows. Each squad dedicated to clearing vulnerabilities is a squad not building product. Backlogs can saturate security and engineering teams, as well as third‑party (often open‑source) maintainers.
Regulatory response
Regulators have noticed. In April 2026 US Treasury Secretary Scott Bessent and then‑Federal Reserve Chair Jerome Powell convened major bank leaders to discuss cybersecurity implications of Anthropic’s Mythos AI model. On July 7, 2026 the European Central Bank’s Banking Supervision asked “significant institutions” to submit action plans by October 31 to accelerate vulnerability and patch management, improve monitoring and AI‑enabled defenses, and strengthen supplier and exposed‑asset safeguards, placing responsibility with management bodies and signaling potential reconsideration of ICT investments and risk‑tolerance frameworks.
The Bank of England, Financial Conduct Authority, HM Treasury, CERT‑EU and national cyber agencies have also issued guidance on frontier‑AI‑enabled cyber risks.
A four‑lever response
Speed alone won’t solve the problem. Effective security organizations need to move four levers together because each depends on the others:
- Cyber practices: AI‑aware detection, rapid containment, stronger identity and certificate controls, multifactor authentication and least‑privilege access.
- Operating model: redesigning how systems are designed, developed, verified and released, and clarifying accountability so recurring flaws are prevented.
- Risk and control framework: shifting from periodic, after‑the‑fact assurance to continuous downside protection—limiting how far an unpatched flaw can reach.
- Technology: platform investments for automation, approved component lists, secure defaults, constrained environments for coding agents, and automated checks in the delivery pipeline.
Practical lessons from early adopters include: different AI models and traditional tools detect largely different sets of flaws, so relying on one tool creates blind spots; preventing recurring vulnerabilities across the organization becomes more valuable in an AI era; and retiring low‑value systems is often the most economical remediation.
How to start: a staged program
Successful organizations begin with a limited scope—selecting a defined set of critical systems and one delivery path—and use the first 90 days to build a working program and evidence for expansion.
- Month 1: define scope, authority and foundations; select critical assets and establish initial technology and governance.
- Month 2: implement the program operationally—automation, prioritization, containment, and remediation processes.
- Month 3: measure material impact, manage ongoing risks and plan scale‑up.
- After one year: the board should assess whether improvements have spread and persisted, focusing on risk tolerances, resource allocation and sustained platform investment.
Conclusion
Frontier AI raises the stakes for software security by increasing the volume of discovery and the pace of exploitation while also supplying tools to defend. Organizations that coordinate investments across cyber practices, operating model, risk and control framework, and technology—backed by clear ownership and funded platform work—can reduce recurring defects, lower material exposure, and improve recoverability. In an environment where flaws are found and weaponized at unprecedented rates, achieving that kind of digital resilience may become a lasting competitive advantage.
Charlie Lewis is a partner in McKinsey’s Connecticut office, Florian Niedermann is a senior partner in the Stuttgart office, Giulio Romanelli is a senior partner in the Milan office, Martin Harrysson is a senior partner in the Bay Area office, Sven Blumberg is a senior partner in the Düsseldorf office, Larry Villella is senior director of digital trust in the London office, and Wolfram Salmanian is a senior expert in the Munich office.
The article was edited by Daniel Eisenberg, an executive editor in the New York office.
