Cybercrime Hit $10.5 Trillion Last Year. AI Defense Is Now the Only Answer

Global cybercrime cost the world an estimated $10.5 trillion in 2025, according to research firm Cybersecurity Ventures — and businesses are now turning to AI to defend against AI-assisted attacks. The arithmetic is stark: that figure exceeds the annual GDP of every country except the US and China. Yet the industry's own answer raises a question it prefers not to dwell on — the same technology breaking the defenses is being asked to fix them.

Unintended real-world breaches executed during training exercises by models from OpenAI and Anthropic have drawn widespread concern, according to the source report. Both companies have restricted access to their most capable models. Experts say it is only a matter of time before less constrained AI systems match those capabilities. Companies are already reporting a rise in cyberattacks, at least partly attributed to AI assistance. The caveat: no one in the industry offers a precise figure for how much of the surge is AI-driven versus AI-detected.

Who is fighting back, and with what?

CrowdStrike is among the cybersecurity firms deploying AI on the defensive side. In September, the company announced SafeMind, a tool built with NVIDIA that deliberately exploits AI's offensive capabilities. It deploys custom models and harnesses to probe simulated versions of customer infrastructure, then fields defensive AI systems to detect the attacks, fend them off, and close any security holes found. Both sides learn from the experience and try again.

"We not only make the changes in the environment before we give it back to the offense and say try again, but we also have a whole manifest—a whole log of every single change that we made to the environment," says Bartley Richardson, chief AI & autonomous systems officer at CrowdStrike. "We give that to the offensive harnesses as well."

"We hear a lot about adversaries and attackers using it, and that's certainly the case," Richardson adds. "What we don't hear as much about is how are we advantaging defenders with the use of AI."

EY is seeing the pressure firsthand. "We are seeing a measurable increase in the volume and velocity of all of these attacks," says David Cooper, cyberstrategy activation leader at the firm. "We see it in our security telemetry clients, security operations centers—we're seeing an uptick." Cooper notes that Microsoft recently unveiled its largest security update in history, and says every company EY works with has deployed AI to help probe and patch its own systems. "There's different levels of adoption, but all of the companies that we work with are using some set of AI tools to help accelerate this."

The World Economic Forum warned earlier this year that "AI is supercharging the cyber arms race." OpenAI, Microsoft, and Anthropic have each launched projects to probe and harden digital systems. Takeaway: AI-versus-AI is now the default posture, not an experiment.

What happens to software nobody maintains?

The problem extends past commercial software into open source — the volunteer-maintained code that underpins much of the internet. Open source projects have reported being inundated with AI-generated bug reports and suggested code changes of varying quality.

"This has been a very substantial burden on upstream, who almost exclusively are unpaid volunteers," says Christopher Robinson, CTO and chief architect of Akrites, an initiative announced by the Linux Foundation and industry partners in June. Akrites coordinates security reports about open source tools and delivers documentation, concrete examples, and sometimes patch code that upstream project leaders can incorporate.

Akrites prioritizes open source projects used in critical infrastructure, including financial services and telecommunications. Where a project has no active maintainer or cannot accept patches, Akrites can release a forked version — a modified copy — and find a maintainer of last resort, Robinson says. The unsaid part: a coordinating body is a patch on a structural problem, not a solution to volunteer burnout.

StackHawk, which tests running applications rather than source code, sees speed as the decisive factor. "The more automation that you add to this process, the faster it goes, and the more things get fixed," says Scott Gerlach, the company's cofounder and chief security officer. "We're just seeing an explosion in rapid remediation, which is something that is foreign, I think, to the world that we typically have lived in," says Joni Klippert, StackHawk's cofounder and CEO.

Most organizations still want a human in the loop before AI makes radical changes to computing environments — both because AI makes mistakes and because humans weigh competing priorities, such as keeping critical systems running. "We cannot really maximize security at any cost," says Mona Rahimi, associate professor of software engineering at Northern Illinois University. "Because we still want to maintain our service."

Is security now a permanent cost center?

Ami Luttwak, cofounder and CTO of Wiz — acquired by Google earlier this year — compares AI's growing hacking capability to a burglar who can enter through a first-floor window one week and reach the second story two weeks later. "Every day, every week, every month, there's new capabilities that AI can do," he says.

Wiz runs its Scan for Good initiative with Google DeepMind, using AI to help critical infrastructure organizations, governments, healthcare institutions, and nonprofits find and fix security holes. Luttwak argues such scans will become routine for all institutions. "This concept of static security controls is dead," he says. "It will be a cat-and-mouse game, and we believe that every company needs to continuously do AI scans for protection."

Whether the current surge of detected flaws slows once AI tools clear the backlog of undiscovered bugs remains unclear. Coding AI tools can introduce bugs of their own, as Robinson points out. For now, the industry's implicit bet is that permanent, automated vigilance costs less than $10.5 trillion a year in damages — a bet no vendor can yet prove.