AI spam filters are getting suckered by old-school text salting

This is less a demand shock than an architecture critique: any security vendor pitching LLM-based email screening as a primary control is now more exposed to procurement pushback, while platforms that combine authentication, URL analysis, rendering, and reputation scoring should look comparatively better. The second-order effect is budget reallocation away from point solutions that over-index on “AI” branding and toward layered stacks where ML is one input among many, which favors broader security suites over narrow email-only products.

The immediate market impact should be modest; buyers already know phishing is adversarial, and this problem has been endemic for years. The more important catalyst is 1-3 months, when enterprise security teams and channel partners start asking vendors for benchmark data on hidden-text handling, false positives, and catch rates; that can pressure pipeline for vendors whose demos rely on content classification alone. Over 6-18 months, expect more spend on provenance, authentication, and browser-based inspection rather than standalone LLM classifiers.

Contrarian view: the consensus may overstate the novelty. The real takeaway is not that AI fails, but that AI used without visibility/context controls fails; that actually reinforces the case for integrated security platforms. If the market sells off “AI cyber” names on this headline, that’s likely an opportunity only if their product is materially dependent on text-only detection and cannot show stronger telemetry-based outcomes. The key falsifier is vendor evidence that hidden-text attacks barely move detection/false-negative metrics in production or that the issue is already solved in recent releases.