# Your Email Filter Is Fighting Last Decade's Phishing
The old tells are gone. No Nigerian prince syntax. No suspicious attachments with .exe buried inside a .zip. The phishing email that landed in your client's inbox this morning was probably indistinguishable from a message their actual colleague sent last Tuesday — same tone, same project references, same casual sign-off. A large language model wrote it in roughly forty seconds.
This is not a hypothetical threat landscape. It's what MSPs are walking into with every new client onboarding, and the math is genuinely alarming: Harvard Business Review put the click-through rate on AI-generated spear phishing campaigns at 54%, matching the performance of human social engineering experts. The difference is cost and scale. A dedicated human attacker running a targeted campaign against a single company is expensive and slow. An AI pipeline scanning LinkedIn and generating a thousand bespoke lures is neither.
## The Four Moves Every AI Phishing Campaign Makes
Understanding the mechanics matters here because it shows exactly where defenses have to shift.
The attack starts with reconnaissance that takes minutes, not hours. An LLM scrapes the target's public footprint — LinkedIn connections, job titles, project mentions, company announcements — and builds a relationship map. It knows who reports to whom, which vendors the company uses, which projects are active. That information becomes the scaffolding for everything that follows.
Content generation is where AI earns its reputation. The resulting email reads like legitimate internal communication because it *is* modeled on it. The spell-check errors and awkward phrasing that made phishing detectable for two decades are gone. What's left is a contextually plausible message that a busy employee has almost no reason to distrust.
Getting through filters is where a technique called polymorphic phishing comes in. Traditional email security relies heavily on signatures — known bad subject lines, sender patterns, URL structures. Polymorphic phishing defeats this by generating a unique variation of every email: different subject line, tweaked formatting, altered link structure. Pair that with delivery through trusted cloud services like SharePoint or Dropbox, and QR codes that redirect through clean intermediary domains, and the gateway sees nothing it recognizes as malicious.
Once someone clicks, the attack escalates fast. Session token theft lets attackers authenticate without needing the actual password. Mailbox rules get created immediately to hide replies and auto-delete warning notices. IBM's 2024 Cost of a Data Breach Report put phishing as the leading cause of breaches — 16% of incidents, averaging $4.8 million per event. The reason those numbers stay high is that post-compromise dwell time is still significant, and it often starts with a single inbox.
## What Filters Were Never Built to Catch
The core problem is architectural. Email security gateways were designed to catch messages that *look* bad. AI-generated phishing doesn't look bad. It looks like your CFO asking about a wire transfer, or your IT vendor following up on a ticket, or HR sending updated benefits information with a link to a portal that spoofs your actual HR system perfectly.
Signature-based detection collapses against polymorphic content. Reputation-based filtering collapses when attackers route through legitimate cloud infrastructure. Content analysis can flag suspicious phrasing, but there's no suspicious phrasing — the email is grammatically impeccable and contextually accurate.
What AI can't disguise is behavioral patterns. The email might look clean, but the *user's subsequent behavior* and the *identity signals around the account* often don't. This is where effective detection now has to live.
## Where MSPs Actually Catch These Attacks
The detection surface has moved. Fighting AI phishing with a better email filter is the wrong battle. The right battle is detecting anomalous identity behavior, endpoint signals, and post-click activity before an attacker has time to establish persistence.
A few specific things that matter:
Identity and authentication telemetry. Impossible travel logins, unfamiliar device fingerprints, first-time geographic access, MFA challenges getting approved from a new browser — these fire even when the initial lure was flawless. Connecting SIEM or MDR tooling to identity providers is now baseline, not advanced.
Mailbox rule monitoring. One of the earliest signs of a compromised account is the creation of rules that filter out security alerts or hide specific sender correspondence. Any new mailbox rule created outside of normal business hours or by a user who has never created one before should generate an alert.
Session token behavior. Credential phishing increasingly targets tokens rather than passwords because tokens bypass MFA. Monitoring for simultaneous sessions from different geographies, rapid context-switches between services, or access patterns that don't match historical usage catches these even after a successful phish.
Endpoint detection at the moment of click. By the time an email filter is relevant, the window has closed. An EDR solution that fires on browser behavior — credential fields on unexpected domains, redirects through known phishing infrastructure, JavaScript execution patterns — catches attacks at the point of impact rather than the point of delivery.
The underlying principle is that no single layer stops this. Email filtering + endpoint detection + identity monitoring + rapid response working together is what stops AI phishing campaigns from turning into $4.8 million incidents.
---
## HackWire Analysis
The uncomfortable reality for MSPs is that the 2024 security stack — a solid email gateway, endpoint protection, and annual phishing training — was already insufficient eighteen months ago. The 54% click-through rate on AI spear phishing isn't just a headline; it means that even with reasonable user awareness, slightly more than half of targeted employees will act on a convincing lure. You cannot train your way out of that math.
What's missing from most coverage of this topic is the economics shift on the attacker side. Polymorphic phishing campaigns that were once cost-prohibitive for all but nation-state actors are now accessible to ransomware affiliates and low-sophistication threat actors who can rent phishing-as-a-service infrastructure. The threat model that justified "we're too small to be targeted" is no longer coherent — if your client has receivables, payroll, or any meaningful data, they are a viable target for automated AI-assisted campaigns that cost almost nothing to run at scale.
The MSPs most exposed right now are those still selling email security as a discrete product rather than as one layer in a behavioral detection architecture. Clients who think they're protected because they have a gateway are operating on a two-generation-old threat model. The conversation with those clients has to happen before the breach, not after. And the framing isn't "your current tool is bad" — it's "the attack changed, and the detection has to change with it." That's a harder conversation, but it's the honest one.
The vendors who figured this out first — the ones selling identity-aware, behavior-driven detection alongside email hygiene — are going to look prescient in the next twelve months as AI phishing matures further.
— HackWire Editorial
---
## Related Coverage