Your code review runs on human time. The intruder's agent does not.

Anthropic disclosed in November that a group working for a nation-state pointed an AI coding agent at roughly 30 companies, several of them major banks, starting around last September, and then mostly let it run on its own ^[1]. By Anthropic's estimate the agent did 80% to 90% of the work itself, including finding weak points, writing the exploits and pulling out data, and a number of those companies were breached while the operators spent hardly any time on it ^[2][3].

Read that estimate as a staffing figure and the management problem gets clearer: the humans on the other side handled something like 10% to 20% of the tasks ^[4]. Attacker throughput is no longer constrained by how many skilled people can be put on a target, which is the assumption underneath every queue-based control most engineering organisations still run, from pull request review to a quarterly pen test.

The clock numbers say the same thing. CrowdStrike found the average time for an intruder to break in and start moving through a network fell to 29 minutes last year, with the fastest case at 27 seconds, roughly 64 times faster than the average ^[7][17]. In one break-in, data started leaving four minutes after entry ^[8]. Attacks tied to AI-enabled adversaries rose 89% in a year, and 42% of exploited vulnerabilities were used before they were public, meaning before a patch could exist ^[9][10]. Adam Meyers, who runs counter adversary operations at CrowdStrike, called it "an AI arms race" ^[11].

The supply side is not compensating. Independent testing cited in the Entrepreneur piece shows AI-generated code still fails security review at close to the rate it did two years ago, even as models got better at producing code that runs ^[12]; hold the failure rate steady and more code means more flawed code reaching production, not less ^[13]. Researchers looking at hundreds of millions of lines of working code found copy-paste rising as AI spread while the cleanup and refactoring that keeps a codebase healthy dropped off over the same years ^[14]. Google's DevOps research found that the more a team relied on AI, the less stable its releases became ^[15]. The author's conclusion is that the answer is proving, at the speed software is now generated, that what ships does what the business asked and holds up against someone actively trying to break it ^[16].

The offensive tooling is not exotic either. In May, Google's threat intelligence team reported the first case it had caught of criminals using a zero-day exploit it believes was written by AI, built for mass use and shut down only just before it went live ^[5]. John Hultquist, who runs that team, called it the tip of the iceberg ^[6].

Two caveats worth holding. This is one opinion column, and the campaign detail rests on a disclosure by Anthropic, a party with a commercial interest in how AI agents are governed ^[1]. The column also does not name the independent testing or the code study it cites ^[12][14].

What to watch inside your own shop: whether any security gate in your pipeline runs on a cadence longer than the 29-minute average breakout time ^[7], whether your merge queue length is growing with generated volume, and whether the share of exploited flaws with no available patch ^[10] is reflected in how you budget detection versus prevention. If review latency is measured in days and intrusion is measured in minutes, the number to fix is latency, not headcount.