This article is part of our June special report on digital risk in family offices.
A new era in software security is dawning. AI models’ coding capabilities are rapidly surpassing humans’, making them dangerously adept at finding and exploiting software flaws.
The turning point was April 7, when AI giant Anthropic announced Claude Mythos Preview—but withheld it from public release due to the risks it poses to the security of software everywhere. During seven weeks of testing, its new general-purpose frontier AI model had autonomously identified thousands of previously unknown “high-severity” vulnerabilities in every major operating system and web browser. Some were decades old.
But that represented only the tip of a titanic cybersecurity iceberg looming on the horizon.
By mid-May, Mythos had surfaced more than 23,000 potential software vulnerabilities across a limited set of more than a thousand open-source systems, with an estimated 6,000 that could be high‑ or critical‑severity flaws.
“The world of cybersecurity has really changed in just a few weeks,” says Alvin Madar, national cybersecurity leader at PwC Canada. “Things that we are seeing today are totally different from what we were talking about even two months ago.”
Software security barriers fall
Cybersecurity experts knew this day would come, but the speed and scale of the emerging threat still come as a shock. AI tools are granting hackers unprecedented capabilities to find weak spots and develop software exploits.
According to the Verizon 2026 Data Breach Investigations Report, 31 per cent of breaches now start with software vulnerabilities—for the first time supplanting phishing for passwords as the most common tactic. The same report found 15 different attack techniques now use generative AI to work faster to spot security gaps and write malware.
“The barrier to entry to hack has gone down significantly,” says Madar. “Now you don’t have to be a very sophisticated technical expert. The threat has become a lot wider because there are so many different AI tools that even unsophisticated hackers can use to launch attacks.”
Call it vibe hacking, the evil twin of vibe coding: opportunistic hackers with minimal skill using AI to thwart cybersecurity measures.
New ways to find old software vulnerabilities
New frontier models like Mythos raise the stakes further, introducing the potential for once rare zero-day vulnerabilities—flaws previously unknown to the software’s developers—to become everyday occurrences.
For instance, some bugs that Mythos found during its testing were decades old, like a 27-year-old vulnerability in OpenBSD, used to run firewalls and other critical infrastructure. Mythos also linked multiple flaws in the Linux kernel, which runs most of the world’s servers, chaining them together to uncover a hidden path to seizing complete control.
“These vulnerabilities are not new vulnerabilities. They have existed forever,” says Madar. “It’s just that humans don’t have the brain power to figure out that these are actual vulnerabilities.”
That’s why, rather than unleashing Mythos on the world, Anthropic instead launched Project Glasswing, restricting the model’s use to roughly 50 major tech infrastructure and cybersecurity firms in order to ferret out flaws and make patches.
“The idea is, don’t expose this vulnerability to the whole world before people have a chance to fix it,” says Robert Moerman, a cybersecurity partner at KPMG in Canada.
It’s not only Anthropic’s models. In early May, OpenAI released GPT-5.5-Cyber, a high-capability cybersecurity variant of its latest model, trained for vulnerability triage, malware analysis and patch validation. Microsoft has its own internal vulnerability-hunting model called M-DASH. And Chinese LLM models are likely catching up, and maybe even surpassing, major Western models.
“I can see all these AI organizations leapfrogging each other,” says Madar. “It’s going to happen very fast. We are not waiting for quarterly releases anymore.”
Update patching strategies
Increasing zero-day vulnerabilities may also make “Patch Tuesday” a thing of the past. Opportunistic hackers will not have to create new vulnerabilities so long as organizations are slow to move on those that go public. “They know that a lot of organizations are not going to be able to patch quickly enough,” says Madar. “They have a window of time where they’re able to exploit these known vulnerabilities.”
Family offices—and their vendors—will need to adjust their patching strategies, making use of automated patching where possible to shrink the risk window, while still navigating the potential for a fix to break an underlying application.
AI is rewriting how organizations need to prioritize patches and address backlogs of vulnerabilities. Gaps that were deemed low risk because only nation states had the tools, techniques or skills to exploit them now need attention. “All those vulnerabilities potentially move from low to critical,” says Moerman. “Organizations that have a backlog of these low- or medium-priority vulnerabilities have to apply human efforts to re-evaluate all of them. That’s going to be a big challenge.”
While Project Glasswing will address the most widespread and critical software exploit risks, there are many other vendors and SaaS providers without protected access to a Mythos-grade AI security vulnerability tool. When such a powerful model does, inevitably, become available, it will be a race for vendors to identify flaws and develop, test and release patches before a hacker finds the same weakness.
Moerman also flags organizations that have written custom applications, often using open-source code not actively looked after or only voluntarily maintained. “When you find a vulnerability in a piece of open-source code written 20 years ago, and it’s now included in everything just kind of by default, who’s going to fix that?” he asks.
Preparing resilience
Madar describes a scenario in which a critical application vendor uses a Mythos-like tool and shares with its customers that it has identified 20 vulnerabilities—but is still working on patches. “At that point, the organization has to decide if they are willing to take the risk of continuing to run the software or take it offline,” says Madar. “And if they do, how do they keep the business operational?”
That’s why he says organizations are folding concerns about growing software exploit risks into business continuity planning. In mid-April, the federal government’s Canadian Centre for Cyber Security launched a new initiative seeking to harden critical infrastructure—energy, telecommunications, transportation and water—in the event of a prolonged cyber disruption. One action item: putting in place plans to isolate systems for up to three months.
Madar says the same thinking needs to apply to family offices. “This is not a technical problem. It’s actually a business process problem that we have to solve,” he says. “We have to ask ourselves, in the event that we are not going to accept certain risks of running an application or using one of the vendors for an extended period of time, how do we keep our organization running and still make money?”
Back to security basics
Moerman argues that the growing spectre of AI-driven software exploits should force organizations, including family offices, to re-invest in their security practices—starting with more oversight into what systems and software are in their environments.
“Everybody’s got end-point detection and response, but almost nobody has very good software or asset inventory,” he says. Without that in place, it’s hard to know what could be vulnerable and what to patch. “Then you’ve got to rethink third-party risk management. What are your vendors’ patch management practices? These are things we should be doing today, but I think the temperature has come up on these issues.”
Questions about cybersecurity practices extend to family office IT service providers. “Some of them do cyber off the side of their desk,” says Madar. “I don’t think you can do it like that anymore, just because of the increased velocity.”
The speed and scale at which AI models can hunt for software vulnerabilities is a wake-up call for family offices. “Toothpaste is out of the tube,” says Moerman. But he is not losing sleep over it—and even sees reason for optimism. “There will be this wave as AI tools demonstrate all these zero-day vulnerabilities,” he says, but over time, software releases will be more secure. “Now software providers are running these models against the code they’re building before they go into production.”
Ultimately, cybersecurity experts preach vigilance, not panic. “You should be aware of it. You should be taking steps against it,” says Moerman. “But we’ll get through this. And to some degree, it is more of the same, just faster.”
Andrew Wahl is a B2B technology content marketing strategist and freelance business journalist. He has written for Canadian Business, Fast Company, BNN Bloomberg, Rogers Business, Bell Business, TD MoneyTalk and BMO Capital Markets, as well as large tech firms in AI, cloud, networking and semiconductor industries. He is based in Hamilton.
The Canadian Family Offices newsletter comes out on Sundays and Wednesdays. If you are interested in stories about Canadian enterprising families, family offices and the professionals who work with them, but like your content aggregated, you can sign up for our free newsletter here.
Please visit here to see information about our standards of journalistic excellence.