Every attack in our dataset had already won. My team analyzed 2,500+ email attacks over the past three months. Every single one had already bypassed Microsoft E3/E5 or a dedicated secure email gateway (SEG)—like Mimecast or Proofpoint—before we detected it. By definition, this is what "sophisticated" looks like in 2025.

The dominant pattern? DocuSign impersonation.

• 281 credential harvesting attacks.
• 13.8% of everything we analyzed.

Here’s the punchline: Pattern matching is failing against these attacks. When we measured how similar the DocuSign attacks were to each other, they averaged just 0.458 on the Jaccard similarity index—meaning each attack shares less than half its features with other variants. A signature that catches one misses most of the others.

Across all attacks, 68% fell below the threshold where pattern matching becomes statistically ineffective. And 38% showed signs of AI-native generation.

This isn’t a tuning problem. It’s an architectural problem. And DocuSign reveals exactly why.

Why DocuSign Is the Perfect Attack Vector

Legitimate DocuSign emails routinely come from strangers, demand urgent action, contain external links you must click, and request sensitive actions like legally binding signatures.

• Traditional View: These are classic phishing indicators. Security awareness training teaches users to be suspicious of exactly these characteristics.
• The Reality: In industries like Law, Finance, and Real Estate, these are also signs of legitimate business.

Attorneys get DocuSign requests from opposing counsel they’ve never emailed. Finance gets signature requests from new vendors. HR sends offer letters to candidates. The urgency is real—court deadlines, closing timelines, board meetings.

This is why document-heavy industries face elevated risk: legal services, pharmaceutical, financial services, real estate. Anywhere high-volume document signing is standard, attackers have a built-in advantage.

Security teams can’t write rules that block malicious DocuSign emails without blocking legitimate business. That’s not a detection problem. That’s a business logic vulnerability.

The False Positive Trap: Damned If You Do, Damned If You Don't

When attacks have low similarity (0.458 index), the instinct is to write broader rules. Loosen the patterns. Require fewer matching conditions.

But broader rules catch legitimate email too. You end up with an impossible choice:

1. Tight Rules: Low false positives, but you miss most AI-generated variants. (Detection Gap)
2. Broad Rules: Catch more variants, but flood your SOC with false positives. (Operational Failure)

One of our customers—a law firm—nearly had an attorney send privileged documents to what appeared to be co-counsel. It was an impersonation attack. Estimated malpractice exposure: $1.5 million. The attack looked exactly like legitimate communication because it was designed to.

Any rule broad enough to catch that variant would flag a significant percentage of legitimate DocuSign traffic. You’d block contracts, delay closings, frustrate users, and generate alert fatigue.

Damned if you do. Damned if you don't.

The Technical Evidence: Why Signatures Are Dead

Let me explain why the similarity numbers matter.

Traditional email security relies on signatures—rules based on shared characteristics. Same sender pattern, similar URL structure, matching content fingerprint. Catch one, catch them all. This works when attacks look alike.

The Jaccard similarity index measures exactly this: what percentage of features do two attacks share?

• Template-based phishing: Typically scores 0.85-0.95. Attackers reuse infrastructure. One signature blocks most of the campaign.
• The 281 DocuSign attacks: Averaged 0.458. Each attack shares less than half its features with other variants.

A signature that perfectly catches one variant misses more than half the others. Security teams can’t write rules fast enough to keep up.

AI Makes This Permanent

38% of the DocuSign attacks showed high AI-assistance indicators.

These aren’t the "Dear Customer, Click Here" templates of five years ago. AI-native attacks reference specific transaction types. They use terminology appropriate to the target’s industry—legal language for law firms, clinical terminology for pharma. They create urgency that feels like real business pressure because it’s modeled on real business communication.

The combination creates a compounding problem: unlimited variant generation at near-zero marginal cost, no consistent signature to match, and faster evolution than signature updates can match.

Current data suggests 35-45% AI-assistance across advanced attack categories. We project 60-75% by the end of 2026. The blind spot is growing.

The Architectural Fix: Reasoning vs. Rules

I’m not writing this to pitch you on a solution. I’m writing because I think the DocuSign data reveals something important about where email security is headed.

When attackers can generate unlimited unique variants, the only stable signals are malicious intent and business context. Not sender patterns. Not URL fingerprints. Not content signatures.

The question isn’t "how do we write better rules?" The question is: "How do we reason about what an email is actually trying to accomplish?"

That’s a fundamentally different detection problem. This is why at StrongestLayer, we focus on building a Reasoning Engine—an Inbox Advisor that sits inside the flow to analyze context, not just code. Legacy architectures built for template-based attacks simply weren't designed to solve this.

The 2,500 attacks in our analysis share one characteristic: they were sophisticated enough to evade the best pattern-matching defenses enterprises currently deploy. That’s not a sample of "all phishing." That’s a sample of what actually causes breaches.

Final Thoughts: The "Silent Failure" of Legacy Security

The most alarming part of this analysis wasn't the 2,500 attacks we found. It was the silence.

These attacks didn't trigger alarms. They didn't flag on a dashboard. They sat in inboxes, indistinguishable from legitimate business, waiting for a tired attorney or a rushed finance manager to make a mistake.

This is the reality of the 0.458 Similarity Index. When attacks are this unique, your current security stack isn't "blocking" them—it is silently validating them. Every time an AI-native phishing email lands in an inbox without a warning banner, your gateway is effectively telling the user: "This is safe."

That is a dangerous lie.

We are moving from an era of Detection (looking for known bad) to an era of Validation (verifying known good). If your defense strategy relies on waiting for a "rule" to catch an AI, you are already breached.

‍

Want the full white paper? Download here.

‍

Frequently Asked Questions (FAQs)

Q1: Why do DocuSign phishing attacks bypass traditional email gateways? 

Traditional gateways (SEGs) rely on pattern matching—looking for known bad links or headers. Our analysis of 2,500 attacks showed that modern AI-native attacks have a low similarity score (0.458 Jaccard index), meaning every attack looks unique. Because they use legitimate links and randomized templates, they slip past rules that look for "known bad" signatures.

Q2: How can I tell if a DocuSign Email is fake if the link is legitimate?

You often can't using standard checks. Attackers now use valid DocuSign infrastructure to host malicious documents, meaning the link is legitimate. This is a business logic vulnerability. Detection requires analyzing the intent and context: Does this relationship exist? Is this transaction expected? This is why a Reasoning Engine is required to catch what rules miss.

Q2: What is the difference between "Rules" and "Reasoning" in email security? 

Rules are binary: "If sender is X, then block." They fail when attacks change. Reasoning is contextual: "Is it normal for this attorney to send a contract at 2 AM to a new vendor?" StrongestLayer’s Inbox Advisor uses reasoning to evaluate the purpose of a communication, not just its metadata, effectively closing the gap where AI phishing thrives.

Q3: Why does tightening security rules lead to more false positives? 

When security teams widen their rules to catch variable AI attacks, they inevitably flag legitimate business emails (like real contracts). This creates false positives that drain SOC productivity. We call this the "False Positive Trap." The only way to escape is to stop relying on broad pattern matching and start using behavioral analysis to distinguish malicious intent from legitimate urgency.

Q4: How common are AI-generated email attacks in 2025?

They are becoming the standard. Our data indicates that 38% of recent sophisticated attacks show signs of AI assistance, with projections reaching 60-75% by 2026. These AI-native threats allow attackers to generate unlimited unique variants at zero cost, rendering traditional signature-based defense obsolete.

‍