As frontier AI models—the most advanced generation of AI systems—make cyberattacks faster and more sophisticated, organizations are increasingly turning to companies such as HackerOne to enhance their defenses. In a conversation with McKinsey Senior Partner Martin Harrysson, HackerOne CEO Kara Sprague explored how AI is changing the risk landscape and why the real challenge is no longer finding vulnerabilities but remediating them before attackers can act.
This interview has been edited for length and clarity.
Martin Harrysson: We are at a moment where AI is reshaping many parts of technology. How did you enter the technology field, and what is structurally different in cybersecurity today compared with before the rise of AI?
Kara Sprague: I started my career at McKinsey, advising clients across the tech stack, then moved into industry at F5, which focuses on application delivery and security. Now I’m CEO of HackerOne, working with some of the world’s biggest companies to find and fix security weaknesses before they can be exploited.
Today, we’re seeing three things happening at once. First, the attack surface is growing because more AI models and systems are being deployed across more enterprises, and AI code generation is accelerating the pace at which companies build and deploy software. Second, that attack surface is more vulnerable because security for AI deployments is still nascent, and AI-generated code still tends to be less secure than human-generated code. And third, attackers are getting more capable because they are quickly adopting and weaponizing frontier AI capabilities.
Defense is behind, and it will take time to catch up. Enterprise security operations will have to retool for speed. And that’s not just about adopting new technology. It’s about rethinking the operating model, talent, and governance. The find-to-fix workflow has to be reimagined with AI at the forefront.
Martin Harrysson: Should leaders see the current landscape as a step change in risk or an acceleration of existing trends?
Kara Sprague: I would call it an acceleration of trends that have been visible for some time. The time between vulnerability discovery and exploitation, for example, has been steadily decreasing from over two years in 2018 to less than a day today.1
What feels new in recent weeks, following Anthropic’s early April announcement of Mythos,2 is the step change in cyber-specific capabilities. Models can now build effective exploits that chain multiple vulnerabilities together, meaning even unsophisticated attackers can identify and target critical issues.
Martin Harrysson: The compression in timelines and the emergence of new attack vectors is significant. How well are organizations keeping up?
Kara Sprague: At HackerOne, we have a broad view across the market because we work with close to 20 percent of the Fortune 500 and almost 40 percent of the Fortune 50. What we’re seeing across our platform is that vulnerability reports have increased 76 percent year on year.3 You might assume those reports are just more noise—what some refer to as “AI slop.” We have seen that in the past, but that’s not what we’re seeing now. The signal—the share of reports that turn into valid findings—has held steady, which means the increase is reflected in real vulnerabilities as well.
Defenders are struggling to keep up with threats. Backlogs are growing. The gap between new threats and the ability to respond is still measured in months, and sometimes years.
It really depends on how quickly security organizations are able to retool their operations for speed. There are emerging blueprints for what that looks like—for example, the Cloud Security Alliance’s recent work on building Mythos-ready security operations4—but closing that gap will require significant changes to how organizations operate.
Martin Harrysson: In practical terms, what steps can organizations take to bridge the defense gap you’re outlining?
Kara Sprague: To meet this surge in AI-led vulnerability discovery, we need to meaningfully accelerate validation and remediation at scale.
At HackerOne, we’re investing in helping organizations overcome a few bottlenecks. First, agentic testing capabilities that incorporate the latest frontier models can find the same vulnerabilities that adversaries using models such as Claude Opus and Mythos would find—but agentic testing does it sooner. Second, agentic validation capabilities will be needed to absorb the coming surge in vulnerability volume while maintaining high signal rates. And third is remediation support—helping organizations move from finding the problem to implementing a verified fix faster and at scale.
Human expertise is critical for implementing these capabilities successfully. At HackerOne, nearly 90 percent of our global community of security researchers uses frontier models to find novel and elusive issues, and they test fixes to ensure they fully remediate exposure.
Martin Harrysson: To your point on the researcher community, how do you see the role of human researchers evolving as AI becomes more capable?
Kara Sprague: Human researchers are shifting from manual bug hunting to higher-level, more strategic work. AI is taking over the discovery of common, high-volume vulnerabilities, while researchers are increasingly focused on building tools and systems that use AI to find deeper, more complex issues such as logic flaws, architectural weaknesses, and novel attack paths.
You can compare autonomous code fixes to self-driving cars: For critical systems, the tolerance for error is essentially zero. Although AI can reliably propose fixes in greenfield projects, it cannot safely untangle complex, legacy architectural issues. Fixing these systems requires expert human oversight.
Ultimately, this evolution will lead to the market placing a much higher premium on identifying complex, high-impact vulnerabilities and guiding critical fixes while decreasing the payouts for lower-complexity issues handled autonomously by AI.
Martin Harrysson: Where do organizations tend to break down as vulnerability discovery accelerates? What does it take to prioritize and act effectively?
Kara Sprague: Organizations tend to break down at the handoff points. The typical workflow—finding an issue, filing a ticket, triaging, prioritizing, and patching—involves multiple handoffs, which create backlogs of work waiting to be processed. Those backlogs don’t hold up under volume. The organizations that perform well are the ones that shorten or minimize those handoffs and ensure findings arrive with enough context.
This is tricky. As the pace of discovery and exploitation accelerates, traditional boundaries between AppSec, DevOps, and SecOps5 become harder to sustain. Separating these functions also creates additional handoffs across teams, slowing response times. Instead, organizations may move toward more integrated workflows with shared data and priorities. Over time, this could mean shifting security from a separate function that inspects software after it is built to something embedded directly into how software is developed and run.
Martin Harrysson: Given those challenges around data, tooling, and organizational alignment, what will separate the organizations that get this right from those that don’t?
Kara Sprague: The biggest difference will be how quickly organizations can clear their backlogs and the extent to which they can compress their find-to-fix workflows. Many are still operating as if they have time between when a vulnerability is discovered and when it is exploited—but with AI, that time is gone.
Organizations that get this right treat it as a time-based problem. Those that fall behind are still managing volume—generating more findings than they can act on. This is where prioritization is important. Traditional vulnerability scoring systems provide a standardized view of severity, but they do not reflect actual risk in a specific environment. The same vulnerability can have very different implications depending on whether an asset is internet-facing, what data it processes, and what controls are in place around it.
Effective prioritization depends on context—combining an understanding of the asset, how a vulnerability can be exploited, and the organization’s capacity to remediate. Without that, even well-identified vulnerabilities can be prioritized incorrectly.
In the end, the advantage won’t go to the organization that finds the most vulnerabilities, but to the one that can reduce its risk exposure the fastest.
Related Articles
AI is the greatest threat—and defense—in cybersecurity today. Here’s why
