AI has cut the cost of a simulated enterprise network break-in by 99% in five months, forcing security budgets to shift from per-seat insurance toward defensive compute. Goldman executives have reportedly discussed spending as much or more on AI for security as on AI coding tools. Security, the article argues, is becoming a computation race: attackers run expected-value calculations on AI tokens, while defenders must convert compute into enough defensive tokens to push that expected value below zero.

In April, Anthropic and OpenAI tested whether Claude Mythos Preview and GPT-5.5 could compromise a simulated British corporate network — a 32-step hack across 20 servers that takes a human expert 20 hours. Both models succeeded without human help, and the Bank for International Settlements valued the break-in at $5k to $10k. Five months later the price had fallen 99%: Claude Mythos 5 lowered it to $770 in June, Claude Opus 5 to $385 in July, and Zhipu‘s open-weight GLM-5.3 matched Mythos’s exploit rate for $90, one-sixtieth of April’s frontier cost. In security operations centers, autonomous investigations compress a 70-minute manual review into minutes for about $2 of compute per alert, helping teams that face more than 4,400 alerts a day and historically lacked the hours to examine two-thirds of them. Still, the cost ratio favors the attacker because defenders must inspect every alert, every hour.

Three multipliers compound defensive compute requirements. Speed: Glean CISO Sunil Agrawal notes AI compresses intrusion response from months to minutes, while CrowdStrike records average breakout times of 29 minutes and a fastest of 27 seconds. Scale: in the July 2026 intrusion at Hugging Face, a swarm of 700 AI agents coordinated over four and a half days, exchanging 70,000 messages and executing 17,600 actions until harvesting 136 production keys; Cogent Research finds 76% of new monthly enterprise attack paths (up 386% year over year) are viable only for parallel AI swarms. Breadth: non-human identities already outnumber employees 82 to 1, and enterprise agent fleets are scaling to 100x headcount, expanding the surface to defend.

The attacker’s math is an expected-value calculation on compute: spending $1m of AI tokens trying to break into a bank with a 1% chance at a $1b payout yields an expected value of $10m and a $9m expected profit, while the same compute on a 5% chance at a $500k payout returns $25k and loses $975k. Since cheap defensive filters block low-effort noise, rational attackers stop spraying small targets and concentrate tokens on the largest balance sheets. The CISO’s job, therefore, is no longer stacking insurance policies but maximizing the efficiency and effectiveness of the security stack: buy enough compute and convert it into enough defensive tokens to drive the attacker’s expected value below zero.