Huskeys' new Profiling beta maps the web endpoints a business depends on

We think the shopper a team pictures when it writes a rule is still a person with a browser. When an AI agent shops on a user's behalf, a security tool working from traffic signals alone has a hard time telling it from an abusive bot ^[10]. In June, Cloudflare's chief executive, Matthew Prince, said more than half of the webpage requests on Cloudflare's network came from automated traffic ^[9]. Huskeys' Profiling launch comes as machines account for more and more of what hits web applications ^[12].

The pitch is a map of what the business depends on ^[1]. Roy Weisfeld, the co-founder and chief technology officer, said that in most organizations everything gets identical protection, "like a mall with one security team guarding every store identically, whether it's a luxury boutique or a snack stand." ^[4] The thing being done is narrower. Once the profile exists, Huskeys judges each request by the standards of the customer's own application ^[3].

Huskeys describes a loop, and the loop includes a check. The company says it verifies afterward whether a change had the effect it was supposed to ^[7]. "The customer decides; Huskeys delivers," said Yotam Mazurik, who leads the company's research team ^[8]. In our view that check is the part a buyer can test, because it should yield a before-and-after number on the buyer's own traffic.

Everything so far is the company's account. Profiling is in beta with a limited set of customers, and Huskeys says results show up within hours ^[11]. The company did not disclose a false-block rate or the number of beta customers. The speed claim covers when results appear. The claim that fewer legitimate customers get blocked is Huskeys' argument ^[16].

We'd run the beta on one path and no more, since Huskeys says customers need no special onboarding to start ^[11]. The tradeoff is reviewer time, because every recommended control needs a customer's sign-off before it goes live ^[6].

To pick the path, sort endpoints on two axes: what a wrongly blocked request costs (high or low), and whether current controls already treat the path differently from the rest (yes or no). High cost with uniform controls is the quadrant Huskeys' pitch addresses, and it is where to start. High cost with controls already tuned needs a comparison against what is in place. Low-cost paths can wait either way.

The forcing function is a count. Record the false-block rate on the chosen path before the beta, and again after one approved change. If the number has not moved, the beta has not yet shown what Huskeys says it does.