Trick or Treat: Can You Tell Good Agents From Bad AI?
Agentic bots now outnumber humans online, but volume isn't the real problem. Here's how to tell legitimate, malicious and chaotic traffic apart at the network edge.
With agentic bots now making up more than half of web traffic, attackers have plenty of cover. The real risk is misclassification: the same request can be a treat for one application and a trick for another, so blanket rules either block good customers and AI agents or let attackers through. Network Edge Security Management (NESM) gives security teams one layer to enforce policy across the CDNs, WAFs and edge services they already run, so every fix fits the application and reaches every edge at once.
Every October 31st, the doorbell rings and you open it to a tiny vampire, a pirate, or a kid in a bedsheet with two eyeholes. You hand over the candy without a second thought. Your website doesn't get that luxury. It opens the door millions of times a day, and the costumes are a lot more convincing.
Who's knocking on your network edge?
Your website doesn't just have visitors anymore. It has customers, crawlers, AI agents, partner integrations, attackers and everything in between. And they all arrive at the same door: your network edge, the CDNs, WAFs, cloud firewalls and edge services standing between the internet and your applications.
Bots now outnumber humans online
According to Imperva's 2026 Bad Bot Report, automated traffic made up more than 53% of all web traffic in 2025. Bad bots alone accounted for 40%.
But volume isn't the real problem. Classification is.
Not every automated visitor is an attacker, and not every useful visitor looks like a human. Your edge has to tell the difference.
From "human or bot" to "legitimate, malicious or chaotic"
For years, the edge had one job: answer "human or bot?" Treat or trick. That model quietly assumed bots were bad and humans were good. Neither holds anymore. (We looked at how agentic traffic rewrote bot defense in Same Game, New Rules.)
The better question is what a visitor means for your business. That puts traffic into three groups.
- Legitimate traffic, the treats: real customers, search engine crawlers, partner integrations and agents acting on behalf of real people. Some of your best visitors now wear a robot costume.
- Malicious traffic, the tricks: credential stuffers, scalpers, scrapers, fake signups and API abuse, often hiding behind residential proxies and headless browsers that fake mouse movements. And not every trick is a bot: some of the costliest fraud comes from real people with real browsers.
- Chaotic traffic, neither trick nor treat: the misconfigured integration retrying in an endless loop. The crawler that bulk-reads your entire site. The script hammering a single endpoint. Nobody broke in, but somebody pays the bill.
{--quote-box--}
Real-world examples of chaotic and legitimate agentic bot traffic
Wikimedia Commons: when crawlers drive up infrastructure costs
Nobody hacked Wikipedia. But since January 2024, the bandwidth used for multimedia downloads from Wikimedia Commons has grown 50%, and the Wikimedia Foundation attributes much of that growth to scraper bots (TechCrunch). Bots generated 35% of the foundation's pageviews, but 65% of its most expensive traffic.
Why? Humans read popular, cached pages. Crawlers bulk-read the long tail, which has to be served from the core data center. In Wikimedia's words, the volume of scraper traffic "is unprecedented and presents growing risks and costs."
Retail: AI-referred shoppers you don't want to block
During the 2025 holiday season, visits to US retail sites referred by generative AI tools grew 693.4% year over year, according to Adobe. These are shoppers who start in an AI assistant and click through. The base is still modest, but it is growing fast, and agents that browse on a shopper's behalf are close behind. A blanket "block anything that looks automated" approach risks turning this demand away.
The lesson: the same traffic pattern can be a treat for one application and a trick for another. The decision has to be made per application, at the edge.
Why your edge security stack can't tell the difference, and your security and revenue being impacted
Most companies already have plenty of locks on the door. The problem is that nobody can see how they work together.
Multi-everything chaos
The network edge is owned by no single team, spread across vendors that don't talk to each other, and until now, impossible to see whole. That's where the haunting starts:
- Shadow rules: nobody remembers writing them, and nobody dares to delete them.
- Missing policies: on the one endpoint attackers found first.
- Blanket rules: they treat every application the same, so you either let tricks in or turn treats away.
- Silent changes: they flip behavior without warning, like when Cloudflare "Log" rules started blocking.
The false positive trap
WAF false positives block real customers at login and checkout, often without anyone noticing until revenue dips. We broke down how it happens in Sorry, Your Order Contains Suspicious Characters.
What is Network Edge Security Management (NESM)?
Network Edge Security Management (NESM) gives teams a way to manage this complexity as one system. It sits above the CDNs, WAFs and edge services they already run, bringing visibility, policy recommendations and orchestration into one layer.
- Assess: the posture across vendors and applications.
- Recommend: policies based on how each application actually behaves.
- Orchestrate: changes consistently across the stack.
NESM in practice
At Hugging Face, this meant 40% fewer unnecessary WAF inspections and over $200K a month in potential mitigation costs prevented. Merlin Entertainment reduced WAF false positives while protecting its global ticketing. More in our case studies.
See who's really knocking: Book a Demo
