I spent years breaking into security systems, and the pattern was consistent. The attacks that worked were always the ones the system had not seen before. New techniques, new infrastructure, and new approaches were the key. 

That’s the dynamic that we’re up against today. Defenders build around what they’ve observed, and then attackers adjust. They don’t have to mimic something legitimate perfectly. It just requires changing the right detail – the one the system doesn’t account for. 

Once they do, there’s nothing to match against. The signal looks unfamiliar, not malicious, and that’s how the attack gets through. 

Why Detection Models Keep Missing Modern Attacks

Email security was built around a different reality. One where the majority of attacks reused infrastructure, relied on known payloads, and followed patterns that repeated across campaigns. Targeted attacks existed, but they weren’t mainstream. Detection systems evolved in response to that structure. Rules and signatures captured known threats, reputation systems tracked infrastructure, and machine learning extended coverage by identifying variations of previously observed behavior.

Even as the technology advanced, the core approach remained consistent. Systems were trained on historical data and optimized to find similarities. That model worked because the majority of attackers continued to produce recognizable signals that could be learned and reused. But these systems were designed before targeted attacks became mainstream. As attacks evolved, those assumptions began to break down, and the most targeted campaigns consistently achieved higher success rates by moving through security layers without triggering detection.

How AI Changed the Nature of Email Threats

AI has removed the constraints that once limited targeted attacks, as it’s able to quickly generate convincing language, replicate context with minimal effort, and apply personalization at scale. What previously required time and precision can now be produced at rapid speed and consistently. The result is a class of attacks that blends directly into legitimate business communication, often leveraging external knowledge and, in some cases, internal data exposed through leaks or compromised vendor and employee accounts, and withstands surface-level inspection.

Where the Signal Actually Lives Now

In this environment, patterns lose their usefulness. The signal has not disappeared, but it has shifted into areas that are harder to evaluate automatically. It now exists in the relationships between participants, in the flow of communication over time, and in whether an interaction makes sense within its broader context.

What Analysts See That Systems Don’t

That shift becomes most visible when looking at how security teams handle these situations. During the early ideation phase of Ocean, we spent time with SOC teams reviewing emails that had bypassed their existing defenses. 

The experience was consistent across teams.  Once analysts investigated these emails, they could identify what the system had missed. The process, however, was long and tedious. In some cases, it took minutes. In others, it required hours or even days.

The Investigation Process

As part of this process, an analyst opens the thread and starts with the basics. They search the sender’s domain on the web, inspect any links, and open attachments or browse to URLs in an isolated environment. These initial checks are meant to surface a quick signal.

If that's not enough, the investigation deepens. The analyst reviews the surrounding context, examines how the sender typically communicates, compares the request with prior interactions, and evaluates whether the timing and content align with expected behavior. They analyze email headers for inconsistencies and trace the message across previous communications.  Within minutes, the investigation surfaces a clear anomaly: the “smoking gun” that reveals the true intent behind the email.

Given the opportunity to investigate, security teams can determine whether a message is malicious with a high degree of confidence. The signal is present, but it exists in dimensions that automated detection systems do not evaluate. This is where the current model reaches its limit.

Why Investigation Has Always Been the Answer

Email investigation approaches the problem from a different starting point. It begins with the assumption that understanding the interaction is required. That means analyzing the content of the message, reconstructing communication history, evaluating participant behavior, and validating whether the request aligns with what is expected in that environment.

This is already how analysts work. They build an understanding of the situation and assess whether it is coherent. That approach holds even when the attack is new, because it does not depend on prior examples and adapts to the structure of the interaction itself.

For a long time, the challenge was applying this level of analysis consistently. Manually investigating every email has never been feasible. Detection systems handled volume, while analysts focused on the subset of messages that required deeper analysis. This created a persistent gap between what could be automatically filtered and what could be fully understood.

Bringing Investigation to Every Email with AI

Agentic AI advancements have changed what is possible. Investigation has always been the most reliable way to understand an email. The limitation was never the method. It was the inability to apply it consistently across the entire environment. That constraint no longer holds because AI can now operate on language, context, and behavior, while also using tools like browsing to URLs, searching the web, file analysis, and parsing content. This allows it to support full investigations, not as isolated checks, but as continuous reasoning processes that evaluate interactions as they unfold.

Each email needs to be treated as an event that requires understanding, which is only possible with an AI-native system that can reason across content, context, and behavior. What previously required manual effort across multiple tools can now be executed continuously and in real time with AI. The result is a conclusion supported by context, rather than a classification based on patterns.

Email attacks today are defined by how effectively they replicate real communication. Attackers operate through context, intent, and timing, while most defenses continue to rely on patterns. That gap is where modern attacks succeed. 

Understanding the email itself has become the only reliable way to secure it. That’s the AI-native approach we’re building at Ocean, combining expertise from social engineering specialists, AI engineers, top-tier analysts who investigate thousands of enterprise email threats, and leading data scientists.

It brings the depth of a full investigation to every email, continuously and in real time.

If you’re still relying on detection as your primary layer, it’s time to see how this model performs in a live environment. Book a demo to experience it firsthand.