IBM prices the AI-assisted breach at a million dollars more than the rest

Divide 2,100 token bundles by 34 hours and the attackers were lifting roughly 62 an hour, overnight, with no shift change ^{[1] [5]}. Spread the same haul across the roughly 200 customer companies reached through one compromised software-as-a-service provider and it comes to about ten bundles each ^{[2] [4]}. Anthropic said AI agents performed nearly all of that work ^[6].

IBM asked 602 companies that had been hacked in the year to February, and about 150 of them said they had been targeted in an attack that used AI ^{[14] [3]}. A quarter of the sample, up 56% on the previous year, implies roughly 16% a year earlier ^{[15] [4]}. IBM put the added damage at an average of $1 million per breach ^[16]. Weight that by the 25% incidence and the expected extra cost on any one breach is about $250,000 ^[5].

The three-hour case is the one that sets staffing. Anthropic's report describes a single stolen developer token becoming administrator control of a victim's cloud environment in about three hours ^[7]. The Google Threat Intelligence Group came at it from the other side, concluding that defenders' response times are shrinking as attackers move beyond piecemeal questions to automating multiple stages of an attack ^[13]. "Sophisticated attacks no longer require sophisticated human attackers," Anthropic said ^[10].

Seoul Economic Daily says the government is rewriting security standards for next-generation AI ^[1], and Anthropic, OpenAI and Google are discussing a joint standards body for the industry, The Information reported on the 13th ^[2]. Two of those three supplied the incident evidence above. Anthropic's material is its own telemetry on misuse of Claude between December and August ^[3]. In the swarm case, one agent parcelled tasks out to sub-agents and ran attacks in parallel against about 50 organisations in education, health care, finance, manufacturing and government. That case was detected and described by the vendor whose model was used ^{[8] [9]}.

I would expect the spending to follow the vendor chain before the perimeter. The compromise of one provider reached about 200 of its clients ^[4], and every one of those clients now has to inventory access it granted and does not run.

Two readings cut against that. The IBM figure is what breached companies said about their own attackers, and attribution follows salience, so part of the 56% rise may be a change in reporting ^{[14] [15]}. And if defensive automation shortens response as fast as offensive automation shortens attack, the $1 million premium is a transitional cost ^[16]. Korea's ministry is careful on this point. Reported breach incidents there totalled 1,236 in the first half, up 19.5% and about 202 more than a year earlier ^{[17] [6]}. "With the spread of generative AI, not every attack is fully automated yet, but when the time and cost of preparing an attack fall, attackers can analyze more targets in less time than before," the ministry said ^[18].

What would show the pricing wrong: a survey next year with AI incidence above 25% and the per-breach premium flat or falling ^{[14] [16]}. That would mean AI is making attacks cheaper to run without making the losses larger.