The Cheaper Attack

Most conversations about AI and cyber attacks start with a scary picture: a superhuman hacker, a perfect phishing email, a machine that finds every weakness. The truth is more ordinary, and in some ways more worrying. AI doesn’t have to make every attack clever. It only has to make parts of an attack cheaper, faster and easier to repeat.

The Realities of AI-Driven Attacks

That’s the pattern emerging from IBM’s 2026 Cost of a Data Breach Report. The global average cost of a breach reached a record $4.99 million, up 12% on the year before. In the United States the figure was $11.5 million, more than double the global average. More than one in four organisations that suffered a malicious attack reported that AI was involved, a 56% increase on the previous year. Those incidents were expensive. AI driven attacks added roughly $1 million to the average cost, taking them past $6 million. Model inversion attacks averaged $6.07 million, and prompt injection attacks came in at $5.89 million.

Look at the reasons and the logic is simple. Attackers spend most of their effort between finding an opening and making use of it, researching targets, writing convincing messages, testing access and moving around. Automate even a slice of that and the economics change. An attack that once needed a skilled person for days can be attempted more often, against more targets, by people who couldn’t have done it before. Leaders tend to underestimate this, and they shouldn’t assume today’s limits will hold.

The Widening Identification Gap

The IBM data also holds an uncomfortable surprise on the defence side. Organisations took an average of 247 days to identify and contain a breach, which reverses several years of gradual improvement. Time costs money. Breaches lasting more than 200 days averaged $5.65 million, compared with $4.32 million for those resolved faster. Teams inside the company identified and contained incidents about 15% faster than external providers, which suggests that knowing your own business pays off when the clock is running.

That brings us to a better question than how to stop every attack. What does the business need in order to keep operating, and can it prove it can recover? Some security leaders call this the minimum viable business. Which services must work? What do they depend on? How long could the company live without them? What would recovery really look like, step by step? The answers belong to operational owners as well as IT. A consumer brand lives on taking orders, fulfilling them and keeping stores open. A business technology company worries about uptime, customer data and showing enterprise buyers it can be trusted. Copying a familiar security roadmap before understanding the business it protects is a common mistake.

Rehearsed Capability vs. Theoretical Plans

There’s an old line in the industry that a plan says what you intend to do, while a test gives you evidence about what you’re able to do. It’s worth taking literally. A recovery plan that has never been exercised is a hope. A recovery that has been rehearsed, timed and fixed after each failure is a capability. Having the right settings on the day of an audit doesn’t answer whether the business survives a disruption.

Used well, AI can shorten that recovery curve for defenders too. IBM found that organisations making extensive use of security AI and automation cut breach costs by $1.93 million and shortened the lifecycle by 65 days. Yet only 36% reported using these tools extensively across the security lifecycle. The gap between the benefit and the adoption is large, and it suggests that many companies are still treating automation as an experiment while attackers treat it as a tool.

Bounded, Reversible, and Observable AI

Where should defenders start? Experienced teams point to jobs where the inputs, permissions and outcomes are well defined: collecting evidence, spotting changes, preparing the first draft of an investigation. Those are the places an agent can help without much risk. Letting a system change production because it sounds confident is a different matter. A useful test is to ask whether an action is bounded, reversible and observable. If it’s all three, automation is a reasonable bet. If it isn’t, a person should decide.

That leads to an uncomfortable point about oversight. Putting a human in the loop only helps if the human has enough context and enough time to challenge the action. Otherwise a button has been added and called oversight. Review needs to get more selective, not more ceremonial. Signing off every tiny step defeats the purpose of automation, but waving through an important change because the queue is long defeats the purpose of review.

Fixing the Unseen Gaps

There’s also a quieter lesson in the 247 day figure. Slow detection often comes from small gaps nobody wants to discuss, a missing owner, an unclear priority, a dependency that nobody mapped. Fixing those isn’t glamorous. Policies, ownership, access and evidence can sound uninteresting, but getting them right makes everything else easier.

None of this means panic. Breach costs are rising, but they respond to preparation. Companies that know what must keep working, can show how they would recover, use automation where it’s safe and keep people accountable for the rest are in a much stronger position than their headline numbers suggest.

The attacker’s job is getting cheaper. The defender’s job is to make sure that cheaper doesn’t mean easier.