Why Bot Detection Has to Stop Chasing Signatures and Start Reading Intent

By Hari Nair, Vice President, Product Management, Cequence Security

Every security team eventually hits the same wall. The signals they built their detection program around turn out to be fakeable. IP addresses rotate in seconds. User-agent strings take one line of code to spoof. Even structural fingerprints, which felt like a real leap forward a few years ago, can be reverse engineered and cloned by a patient attacker within minutes. Once a team realizes this, the natural next question is what actually holds up. And the answer the industry keeps arriving at is whether through internal research, red team findings, or hard operational evidence, is behavioral intent. 

At first, that distinction sounds subtle until you realize the scale attackers now operate. One campaign against a single organization generated nearly 300,000 distinct behavioral profiles in a 24 hour window. That works out to a new variant roughly every 300 milliseconds. The attacker kept changing every characteristic they could, but their goal never changed. Today, effective bot defense comes from identifying what remains consistent beneath all that change.

The blunt instruments era

Before bot traffic became a board-level risk, most defenses were blunt by necessity. IP-based blocking worked until an attacker rotated addresses, which took no effort at all. User-agent matching was even more fragile, since a single string substitution defeated it. Signature-based approaches from WAF and CDN vendors held up against known patterns but fell apart the moment something novel appeared. However, none of these approaches looked at how a request behaved in the context of the session around it. They only looked at the request in isolation and asked whether it matched a known bad shape.

Structural fingerprinting was the next real advance, and it’s the approach our team built early on. Every browser request carries small structural details, like the order its headers appear in, which fields are present or missing, and how it encodes language and compression preferences. Our system captured those details and compressed them into a single fingerprint. That gave defenders something far more useful than an IP address to write policy against. Bots could mimic browser traffic closely, but they still tended to produce a structural fingerprint that looked different from a real session hitting the same endpoint. For a while, that difference was enough to catch them.

Why a fixed target always loses

Attackers adapt, and any fixed algorithm presents a fixed target. The hardest problem to solve turned out to be the “mixed” fingerprint, the case where attacker traffic and legitimate traffic hash to the same value because a bot has cloned enough structural characteristics of a real browser. At that point, blocking on the fingerprint creates risk in both directions. When acted on aggressively, real customers get caught in the net. If held back, the attack goes through uncontested.

Unsurprisingly, sophisticated operators found this gap quickly. A replay attack is simple to build: capture the structural characteristics of a genuine browser session, then replay those requests from a script. The result? Traffic looks statistically identical to a real user, and there is no signal left to act on.

This core lesson is exactly what the industry has converged on. A one-size-fits-all algorithm, applying the same fixed feature set across every customer, endpoint, and traffic pattern, has no way to adapt once an adversary has studied what that algorithm actually examines. The fix isn’t a better algorithm. It’s a different architecture entirely and one built to read intent rather than shape.

What reading intent actually requires

Instead, it looks at several different behavioral signals at the same time, and each one captures a different piece of what an entity is doing. Put together, they build a fuller picture than any single signal could on its own. If an attacker changes one piece of their behavior to slip past detection, the other signals still catch what they’re actually trying to do.

Two recent incidents show why that matters:

• In the first, an attacker targeting a connected-device platform ran into blocking policies built on this behavioral approach. A developer community maintaining the attack tooling began rotating identifiers and restructuring requests in real time, probing for any edge that wasn’t covered. A static fingerprint would have lost this race. Here, each new variant still got caught, because rotating one piece of behavior didn’t change the intent behind it. The customer’s reaction when the policy went live was blunt: “I’m ecstatic.”
• The second was about volume rather than adaptation. A campaign against a major consumer-facing brand generated 297,556 distinct behavioral profiles in a single day, one new variant roughly every 300 milliseconds. A defense chasing individual signatures would have stayed permanently behind that pace. What held up was focusing on the behavioral invariants, the parts of the attack that couldn’t change without breaking its own function. Those fed straight into mitigation policy, and the attack was contained.

Neither of these incidents were won by a single clever rule. In both cases, they were contained by an architecture that could compose the right behavioral signals for that specific threat, and then acted on them without requiring a human to sign off on every decision. Security teams evaluating bot and fraud defenses today need to be asking vendors (and their own programs) if detection is built around identifying a fixed shape, or around modeling what an entity is trying to do? The first approach has an expiration date the moment an attacker studies it closely enough. The second one doesn’t, because intent is a much harder thing to rotate away than an IP address or a header order.

This distinction is about to matter more as the agentic AI expands the attack surface. We can expect to see more automated entities interacting with APIs and applications at machine speed, with behavioral signatures that are often more consistent, not less, than anything a human user produces. That consistency is a gift to any framework built to read intent, and a serious liability to any defense still built around matching a static signature. The organizations that get ahead of this shift will be the ones that stop asking what a request looks like and start asking what it’s actually trying to do.

About The Author

Hari Nair leads product management at Cequence Security, working closely with prospects, customers, and internal users to define the product strategy and roadmap that keeps the Cequence products at the cutting edge of API, bot, and AI security.

Before joining Cequence, Hari led IoT security at Aeris, managed Cloud Security at Open Raven, and drove Machine Identity at Venafi.

He holds degrees in Engineering from Trivandrum, Management Science from Stanford, and is CISO-certified by Carnegie Mellon.