LLM detection, reads the code, reasons about intent.

Every signature-based scanner is answering one question: have we seen this exact thing before? When the answer is yes, it is the best tool available, fast, cheap and essentially never wrong. When the answer is no, it has nothing to say.

That limitation used to be tolerable, because writing new malware took effort and campaigns reused code for months. It is much less tolerable now. Generating a functionally identical backdoor with a different shape is close to free, which means the gap between a novel sample and a catalogued one has widened into the place where most successful compromises live.

This is the specific situation people describe when they say four scanners called the site clean and it was not. The scanners were not broken. They were asked whether the code matched something known, and it honestly did not.

The LLM layer asks a different question: what does this code do, and is that something a legitimate plugin would need to do? A function that reads a value from the request, decodes it, and executes the result is a remote shell. That is true regardless of what the variables are called, how many layers of encoding it hides behind, whether it has ever been catalogued, or how convincingly it has been made to look like part of a caching library.

The same reasoning also reduces false positives, which is the half of the story heuristic scanners struggle with. A minifier's output is dense and unreadable but has no interest in your request parameters. An obfuscated licence check phones a vendor's server and validates a key. Both look alarming to a heuristic and neither is a backdoor, and a layer that reasons about purpose can say so.

Every finding comes with an explanation written for a person rather than an analyst: what the code does, why it was flagged, what an attacker would use it for, and what to do next. A file path and a CVSS score tell an experienced developer plenty and tell a site owner nothing, and at nine on a Sunday evening the person deciding is usually the site owner.

The AI layer is what Pro is for. Everything beneath it (signature, heuristic, integrity and database scanning) is in the free plugin, because a scanner that cannot find catalogued malware without payment is not a scanner. Reasoning about intent is the expensive part, computationally and otherwise, and it is the part that handles what nothing else can.

It is not magic and it is not a replacement for the other layers. It runs on top of them, on the code they surface, which is why the free layers are not degraded to make Pro look better. A signature match is still the fastest correct answer available, and when one is available the AI layer is not the thing that found your problem.

It also does not remove the need for judgement on a genuinely ambiguous file. What it does is put the ambiguity in front of you in language you can act on, rather than leaving it silently unreported because it matched nothing.