Kurtz frames his first question about agents around location: where they are, and his answer is that you cannot even find them, followed by whether they are touching data or connecting to systems they should not be [9]. Anyone who has run an asset inventory knows that sequence. Discovery is the problem the security team lives inside; model quality is the problem that gets the keynote.
The two numbers in the pitch are worth holding apart. There is 52% better at identifying vulnerabilities and issues than the best frontier model, on CrowdStrike's test [7], and there is 98% cheaper to run [8]. The cost figure has the operational teeth: at 2% of the price, one frontier run's budget buys roughly fifty SafeMind runs [12]. A cheaper finder that surfaces more findings simply moves the work downstream, piling the queue onto whoever triages. That is where the autonomy ladder starts carrying weight. Kurtz laid out five levels of SOC autonomy last year and says he is now targeting the top of that scale, with the Cyber Superintelligence Lab built to reach the point where the car drives itself and the human is kept in the loop only for critical calls [10]. Full autonomy is still a roadmap goal; the release notes describe something less.
What the SiliconANGLE account does not carry: which frontier model was the comparison, what the test set contained, or whether "52% better" means more true findings or fewer false ones [13]. Those details are the difference between a benchmark and a demo.
Now separate the thing being pitched from the thing being bought. Pitched: a two-model loop, one side finding the attack path and the other closing it, harnesses driving both [5]. Bought, for most organisations: coverage. Kurtz's argument is that AI gets consumed on endpoints and in cloud containers, ground CrowdStrike already sits on, and that the Signal AI acquisition pulls the same visibility into identity [4]. CrowdStrike also says it is measuring its own models against general-purpose ones and offering customers both [6], which is a fair position and hands the choice back to the buyer.
Here is the grid worth drawing before the demo. One axis: can you enumerate the agents running in your environment this week. The other: can your SOC close a finding without a human approving each step. Bottom left, no inventory and no autonomy, the model comparison is beside the point and what you are actually buying is discovery. Top left, inventory but no autonomy, cheap inference works against you, because fifty times the runs [12] means generating findings faster than analysts clear them. Bottom right is automating action across a population you cannot list. Only top right turns a 52% accuracy edge [7] into shorter dwell time.
Most teams sit bottom left and describe themselves as top left. The figure worth asking a vendor for is not the accuracy delta but the share of findings that closed last quarter with no human step in the path, because that is what tells you which quadrant the purchase lands in.
