The math finally broke in the attacker's favor
For years, security people said something like "attackers only need to be right once, defenders need to be right every time" and treated it as a philosophical nicety. In 2026, it became a spreadsheet problem. IBM's newly released Cost of a Data Breach Report, based on 602 organizations hit between March 2025 and February 2026, found that one in four malicious breaches were AI-enabled. That's up 56% from the year before. And those breaches averaged $6 million each, about a million dollars more than the overall breach average of $4.99 million.
The reason isn't mysterious. Launching an AI-assisted attack is getting absurdly cheap, in some cases costing attackers only thousands of dollars to find and exploit new flaws within hours, according to reporting tied to Anthropic's own red-team findings on frontier models like Mythos. Meanwhile, cleaning up after a breach still takes months of forensics, legal review, customer notification, and regulatory paperwork. Attack cost is falling. Response cost isn't. That gap is the whole story.
Deepfakes and malware, not sci-fi autonomous hacking
It's worth being precise about what "AI-enabled breach" actually means here, because it's less Terminator and more mundane fraud at scale. IBM found these incidents were mostly deepfake impersonation and AI-generated malware, the kind of thing that lets a mid-tier criminal crew suddenly sound and write like a company's CFO, or churn out malware variants faster than signature-based tools can catalog them.
Critical infrastructure took the brunt of it. 62% of AI-driven attacks in the report targeted critical infrastructure sectors, with financial services breaches averaging $6.3 million and energy sector breaches averaging $5.2 million. That concentration matters beyond the individual company. When AI attacks cluster around banks and power grids, a single successful campaign can ripple into supply chains and other essential services, not just one victim's balance sheet.
Companies know AI attacks AI, but adoption is lopsided
Here's the part that should sting a little. IBM found that organizations using AI and automation in their security operations cut breach costs by nearly $2 million on average. The tools work. And yet one in four organizations still haven't adopted them in security operations at all.
Even among the companies that have, the deployment is uneven in a telling way. More than half are using AI agents for threat detection and containment, the flashy, dashboard-friendly stuff. But only 18% apply agents to vulnerability management, which is the unglamorous work of actually patching known holes before someone walks through them. That's backwards, given that attackers are moving faster specifically because unpatched, known exposures are still sitting there waiting. IBM's Suja Viswesan put it plainly: the priority now is closing the lag between discovery and remediation, not just getting better alerts.
What's actually changing in response
The more interesting shift isn't in the breach numbers, it's in when companies decide to act. In a Ponemon Institute follow-up conducted after the initial breach research, 85% of organizations said they planned to increase security spending simply after learning about frontier AI cyber capabilities, compared to 64% who said they'd increase spending only after experiencing an actual breach. For the first time, more companies are reacting to the threat of capability than to direct pain. Three quarters said frontier AI risk is prompting them to rethink how they deploy their own defensive agents.
There's also a quieter weak point buried in the report: 20% of organizations reported a breach targeting AI models or applications directly, and the most common causes weren't exotic model attacks at all. They were compromised APIs, plug-ins, and cloud misconfigurations sitting around the AI systems, each cited in 27% of cases. In other words, the AI itself often isn't the hole. The ordinary infrastructure wrapped around it is.
None of this means the fight is unwinnable, but it does mean the old cadence, patch on a schedule, review access quarterly, train staff once a year, is built for a slower opponent than the one showing up now. The organizations pulling ahead aren't the ones with the flashiest AI security tool. They're the ones who closed the gap between "we found the problem" and "we fixed the problem," because that gap is exactly where the $6 million number lives.
Thumbnail: Unsplash, License: Unsplash License (free to use)