It's counterintuitive, sure — but the numbers don't lie: phishing volume dropped nearly 40% between 2024 and 2025, according to Zscaler's annual phishing report.
And yet, the threat? Higher than ever.
Why does fewer emails feel riskier than more?
Because the ones that do get through today aren't spamming you — they're hunting you.
Hackers have quietly shifted from spray-and-pray to surgical strikes, and AI is their scalpel. Instead of firing off a thousand generic offers for $50 iPhones from "[email protected]", threat actors are now crafting bespoke emails designed to look like they came from your actual payroll team, your bank's fraud department, or even your boss — complete with the right tone, the right timing, and the right hook. The quality of each message has leapt forward; the quantity, deliberately curtailed. The math changed: fewer emails, but each one packs a much bigger punch.
This isn't just about clever copywriting. The infrastructure, the personalization, and yes — the AI underneath it all — has gotten way more disciplined. And for defenders? That means the old "delete and hope" instinct no longer cuts it.
AI's Double-Edged Sword
When ChatGPT dropped in late 2022, the obvious assumption was that phishing would surge. Why? Because generative AI could instantly turn anyone with a decent prompt into a capable copywriter, with access to phishing kits that would auto-generate lures, images, landing pages, and even evasion tactics.
Zscaler actually saw exactly that surge — a 58% jump right after ChatGPT's launch. But the curve reversed hard soon after. By 2024, volume dropped 20%, and another 20% in 2025. That's not a cooldown — it's a strategic pivot.
Brett Stone-Gross, Zscaler's senior director of threat intelligence, sums it up well: "Instead of going en masse, they're doing more targeted attacks. That requires more effort and resources, but the payoff is better."
Think of it like street crime turning into home invasion. Why mug five people and walk away with $10 each when you can pick one target, research their habits, and walk out with $10,000? The same logic applies here: the ROI on targeted phishing is simply higher — and AI helps you find those targets and tailor the hook without bloating your volume.
Here's what changed:
- Lower effort per targeted email: Early AI lures often sounded robotic and needed a lot of proofreading. Today's models — fine-tuned on legitimate corporate comms, invoices, and HR templates — produce eerily plausible drafts in seconds.
- Better domain impersonation: AI isn't just writing the email. Some phishing kits use AI to generate fake domains that look like trusted brands without actually registering them. Think
apple-support-login[.]comvs. the realapple.com. Generators can spin these up with minimal manual cleanup. - Campaign orchestration: AI now suggests when to send — when your target is most likely to be awake, logged in, and not busy on Zoom. Timing matters more than volume.
And when those targeted emails do land, the payoff multiplies — especially if attackers combine them with MFA bypass techniques like device code phishing, turning a single credential into full account takeover.
Numbers Don't Lie — But They Do Mislead
The FBI's 2025 Internet Crime Report tells a stark story. In 2024, phishing complaints and losses looked manageable — $70 million total. Same number of complaints in 2025? Almost. But losses tripled to $215 million.
In 2023, the FBI received more complaints than either year, yet losses totaled just $18 million.
What gives? It means each phishing email today isn't dragging in a few extra dollars. It's extracting hundreds or thousands per victim.
Zscaler's data backs this up: while education sector phishing dropped 66% in 2025, services attacks rose 66%, and government sector attempts went up 50%. Why those? Because they're high-value targets — think payroll systems, wire instructions, or procurement portals. And attackers are finally aligning their tools with that intent.
And the geographic angle? Phishing infrastructure shifted dramatically, too. Brazilian hosting rose 2,522%; Hong Kong dropped 90%. That's not chaos — it's strategy. Attackers aren't just spamming from random IP ranges anymore. They're curating their hosting stack to avoid takedowns, minimize latency, and blend into normal traffic.
Cloud Hosting — Where Attackers Hide in Plain Sight
Here's a wild stat: Of all the attacker IPs that Zscaler's decoys detected, 76% came from Amazon Web Services (AWS) address space.
That's not because AWS is bad. It's because it's almost too good for attackers.
Stone-Gross explains: "One is cost — AWS instances are quite cheap. And the other is: I think Amazon's abuse department is probably overwhelmed."
Let that sink in. Attackers aren't building custom infrastructure or bouncing around bulletproof hosts. They're spinning up a t3.micro on AWS because it's cheap, reliable, and hard to distinguish from legitimate traffic.
Here's why AWS is so attractive:
- Cost-effective: You barely pay a cent to run an HTTP server that tricks someone into entering credentials.
- Global reliability: Downtime kills phishing campaigns. AWS offers 99.9% uptime SLAs — better than most shady bulletproof hosts.
- Evasion: Blocking AWS IP ranges is nearly impossible — you'd block half the internet. Attackers know this and lean into it.
And here's a chilling side effect: legacy blocklists are now almost useless. If every email comes from a legitimate-looking IP (like AWS), then simple IP filtering buys you little. You need behavioral monitoring — or as Stone-Gross puts it, "more specific information." An IP address is no longer specific. It's just a starting point.
The Talent Shift — Smaller Teams, Sharper Tools
The old narrative was "bigger operations = more phishing." Not anymore. Today's most dangerous campaigns are run by lean, focused teams — often just two or three people with a single AI-powered toolkit.
This has real implications:
- Specialization matters: One person might specialize in domain mimicry; another in UI redressing for fake login pages. AI lets them coordinate rapidly — sharing templates, tactics, and even live feedback on open rate metrics.
- Low barrier, high ceiling: A single motivated actor with decent prompting skills can launch a multi-stage phishing campaign now. That wasn't true five years ago.
- Rapid iteration: AI tools let attackers A/B test subject lines, payload timing, and even visual cues in near real time. The feedback loop is minutes instead of weeks.
Effectively, AI isn't just automating phishing — it's democratizing sophistication.
The result? A new breed of phisher: not a script-kiddie with a botnet, but an operator who treats email like marketing — research, audience segmentation, A/B testing, and conversion optimization. The only difference is they're harvesting credentials instead of clicks.
What Defenders Can Actually Do (Hint: It's Not Blocking)
Traditional network security — IP blocklists, signature-based detection, static filters — is quickly becoming a relic.
AI-powered phishing is too context-aware for blunt tools. So what works?
- Behavioral baselining: Know when this user normally logs in, what their timezone is, and whether an attachment is out of character. If a user rarely shares files but suddenly receives a "Urgent: sign this W-9" from their own boss, that's a red flag. Not because the sender address is fake (it usually isn't), but because the behavior doesn't fit. The industry is already pivoting toward behavioral AI email protections that detect anomalies instead of matching signatures.
- Deep email authentication: DMARC, DKIM, and SPF are table stakes now. Look beyond them: do the sender's metadata fields actually match? Do the reply-to and from addresses align, or is one a spoofed alias? Even with these in place, misconfigurations can leave gaps — as explored in how Exchange "sender" field spoofing enables active attacks.
- Human-in-the-loop detection: AI defenders can help flag anomalies, but a quick voice call to the sender — "Hey Sam, did you just email about wire transfer instructions?" — is still the gold standard. Leverage your team's social context, not just their technical savvy.
And maybe most importantly — stop treating users as the weakest link. Train them to question urgency, but also give them a frictionless way to report suspicious messages without fear of blame.
The truth is, if an attacker can make a phishing email look just convincing enough to bypass your network — and then make it feel emotionally urgent — you need more than filters. You need context. You need judgment.
AI gives attackers tools to raise the bar. Your job now isn't to lower it; it's to raise it higher.
Final Word: The Risk Is Higher. So's the Reward.
The phishing threat isn't going away — it's getting smarter.
With volume down and quality up, defenders can no longer rely on quantity-based heuristics. You can't block what looks legitimate from the outside but reeks of malice on the inside.
The real lesson? Don't just scan for obvious red flags. Scan for inconsistencies — in tone, timing, wording, or behavior. That's where today's AI-augmented attacks live: in the tiny gaps between "real" and "just realistic enough."
And if you're still teaching users to spot phishing by looking for typos? You're already behind.
The bar has been raised. Time to clear it.