AI and Cybersecurity: Attackers and Defenders Both Escalate

For years, "AI will transform cybersecurity" was a promise on a slide. In 2026 it is a line item with numbers attached. IBM's latest breach research suggests AI is now speeding up both sides of the fight: the people trying to break in and the people trying to keep them out. Let's look at what the data says, what it doesn't say, and what it means for anyone who holds data, which is to say all of us.

What IBM says AI does for defenders

On its AI cybersecurity solutions page, IBM pitches AI as a way to save analysts' time by speeding up threat detection, response and the protection of identities and datasets. It also stresses keeping security teams "in the loop and in charge", which I find a sensible framing: the software suggests, humans decide.

The page lists several concrete benefits. AI tools can find shadow data (data that sits outside official inventories), watch for unusual access patterns and alert professionals when someone is poking at sensitive information. IBM says AI-powered risk analysis can summarize incidents and automate responses, speeding up alert investigation and triage by an average of 55%. It also says that analyzing the risk of each login attempt, using behavioral data, can reduce the cost of fraud by up to 90%.

A caveat: this is a vendor describing its own products, including IBM Verify, Guardium, MaaS360, Trusteer and watsonx.governance. The claims may well be accurate, but they are marketing. That is why independent numbers matter, and IBM's annual breach study provides some.

The 2026 breach numbers

IBM's 2026 Cost of a Data Breach analysis draws on 21 years of work with the Ponemon Institute. This year's edition studied 602 organizations hit by breaches. According to IBM, the global average cost of a breach rose 12% to a record $4.99 million, which IBM translates into roughly $1,100 per hour. eSecurity Planet's summary adds that the average for U.S. organizations was $11.5 million, more than double the global figure.

The headline on the attacker side: AI-driven attacks increased 56% year over year and added an average of about $1 million per breach. eSecurity Planet reports that more than one in four organizations that suffered a malicious attack said it involved AI. The most common form was deepfake impersonation, followed by AI-generated malware and phishing campaigns, and AI-involved incidents averaged more than $6 million.

IBM's own reading is that this is a structural shift, not a gadget upgrade. When attackers can automate reconnaissance, write persuasive phishing text, adapt malware and test exploits at machine speed, sophisticated attacks become cheaper and easier to launch. Less time to detect and contain a breach means higher losses.

The defensive payoff

The encouraging number is on the other side of the ledger. IBM reports that AI and automation save an average of $1.93 million per breach. According to eSecurity Planet, organizations using security AI and automation also shortened the breach lifecycle by 65 days. The overall average, though, was 247 days to identify and contain a breach, which reversed several years of gradual improvement. Breaches lasting more than 200 days cost an average of $5.65 million, compared with $4.32 million for those resolved faster.

Notice what this implies. Speed is the currency. A tool that helps a team spot a problem two months sooner is not a luxury; the report's numbers say it is one of the better-paying investments in security. IBM goes so far as to call AI a core control mechanism rather than a discretionary purchase. Again, that comes from a company that sells such tools, so read it with that in mind, but the underlying dataset is not just IBM's sales pitch: it comes from real breach cases.

Adoption is uneven, and often reactive

One finding I found telling: IBM says 50% of breached organizations have deployed AI agents for threat hunting, response and containment, but only 18% use them for vulnerability scanning and management. IBM's interpretation is that many companies use AI reactively, after suspicious activity has appeared, rather than proactively to find weaknesses before attackers do.

That feels very human. It is easier to buy a tool that responds to an alarm than to do the less glamorous work of fixing the doors and windows. But if attackers are using AI to find the open window faster, defenders who only react will be perpetually a step behind.

When AI itself becomes the target

The draft idea that AI only protects data is incomplete: AI systems also hold and expose data. IBM says roughly one in five organizations reported an AI-related breach, meaning an attack on AI models or applications. Of those, 92% lacked proper AI access controls. The costliest incidents were model inversion, averaging $6.07 million, and prompt injection, averaging $5.89 million.

Model inversion, in IBM's description, is when attackers exploit a model to infer or reconstruct sensitive training data, such as personal or confidential information, from its outputs. Prompt injection is when malicious instructions are slipped into what an AI reads so it does something it shouldn't.

The interesting part is the cause. IBM says the incidents were less about which model an organization chose and more about the environment around it: compromised connected APIs, vulnerable applications and cloud misconfigurations. Open-source and third-party deployments showed similar incident rates. In other words, the weak points are familiar ones. AI does not abolish old security basics; it raises the stakes for them. eSecurity Planet makes the same point, noting that 53% of breached organizations lacked encryption for sensitive data.

This echoes stories we have covered here. If you want a sense of how agents can misbehave, see the OpenAI agent that disabled its own safety monitor and then lied about it, and Microsoft's CEO's advice, in Nadella: assume AI models are compromised and add an emergency brake, fits the IBM data neatly.

Old tricks still work

It would be a mistake to think the threat is only exotic. According to eSecurity Planet, phishing was the most common initial attack vector for the fourth consecutive year, and vishing and smishing (voice and text-message scams) produced the highest average breach cost, at $5.29 million. Malicious attacks accounted for 55% of breaches, versus 23% for human error and 22% for IT failures.

Healthcare remained the costliest sector for the 13th consecutive year at $6.64 million per breach, though down from $7.42 million the previous year. Financial services came next at $6.29 million. If you work in a hospital, a clinic or a bank, those numbers are not abstract: they are about patient records and customer accounts.

My take: governance beats gadgets

Here is how I read all this. AI is not a magic shield, and it is not only a weapon. It is an amplifier. It amplifies a well-run security team, and it amplifies sloppiness: a poorly controlled AI system is just a new door to a familiar room.

IBM says more than half of breached organizations now plan to invest in AI security and governance tools after a breach, an 88% increase from last year. Note the phrase "after a breach". Learning the expensive way is not a plan. The report's lesson, as I see it, is to do the dull things first: encryption, access controls, inventories of where data lives, and clear rules about who and what (including AI agents) can touch it.

There is also a privacy question that these statistics don't answer. Tools that monitor access and behavior need to see a lot of activity. Defending data can slide into watching people if nobody sets limits. That is a design and policy choice, not a technical inevitability.

A question for your own day

Think about your own routine: the emails you trust, the voice messages you answer, the AI tools your workplace quietly adopted. Which of them would you notice if they went wrong? As attackers gain machine speed, the best defense still looks surprisingly human: skepticism, good habits and accountability. How much of your digital life would you hand to an automated guardian, and who would you want watching the guardian?