Interviews
Oran Moyal, Co-Founder and CTO of Ocean Security – Interview Series

Key PointsAI-native status is organizational rather than technical because it changes how every…
View more
- AI-native status is organizational rather than technical because it changes how every team handles work.
- Speeding individual tasks fails when handoffs and approvals remain the real bottlenecks between steps.
- Agents need enforced permissions and independent validation because instructions alone are not security controls.
Oran Moyal is Co-Founder and CTO of Ocean Security, where he develops agentic AI for investigating sophisticated email threats. His background spans cybersecurity research and development, offensive social engineering, and enterprise security. He co-founded Ocean in 2024 with Shay Shwartz, drawing on their experience analyzing how attackers exploit trusted communication and how security systems can respond to increasingly targeted threats.
Ocean Security builds an agentic email security platform designed to investigate targeted phishing, business email compromise, and financial fraud. Its specialized AI agents analyze communication context and sender intent to help security teams identify threats within seemingly legitimate business interactions. In this interview, Moyal discusses what makes an organization AI-native, how to redesign workflows around agents, and why clear requirements, independent validation, enforced permissions, and human judgment remain central to effective automation.
Before co-founding Ocean in 2024, you spent years working in cybersecurity research and development, including roles in the Israel Defense Forces and at VisibleRisk / Microsoft Security. What experiences led you to launch Ocean, and how did the emergence of AI influence your vision for the company from the outset?
A targeted spear phishing campaign used to take a dedicated team and weeks of manual research and preparation. When generative AI arrived, we realized that every employee in every enterprise is now subject to a devastating nation-state-level phishing attack. Overnight, a $20 subscription gave every attacker nation-state capabilities.
Shay and I have deep social engineering expertise from the offensive side, and we’ve both built and scaled enterprise security companies, so we knew legacy tools couldn’t keep up. They look for rule violations and anomalies, but these attacks live inside normal business communication: real threads, trusted accounts and legitimate services. Nothing looks abnormal. So, in 2024 we started Ocean to help enterprises adapt to this new threat with AI-driven email security that analyzes context and sender intent in real time, much like a human investigator would, before a compromise occurs.
Many organizations describe themselves as “AI-native” simply because they’ve adopted AI tools or automated certain processes. In your view, what fundamentally distinguishes a truly AI-native organization from one that has simply integrated AI into its existing operations?
Many companies call themselves AI-native because they’ve bought AI tools. That’s technology adoption. Being AI-native changes how the company handles everyday work and how it solves bottlenecks.
It starts with how leaders respond to a bottleneck. The traditional reflex is “This takes too much time; let’s hire someone to handle it.” An AI-native organization first asks, “How can AI speed up this entire function?” Repeat that question across every team, and the whole company starts operating differently.
Real AI-native transformation is organizational, not technical. If you speed up one department, the bottleneck just moves to the next one, and the company ends up moving at the pace of its slowest function. AI-native organizations change together, so information flows faster, feedback loops get shorter and teams work as one connected system.
In this model, the value of people shifts. As AI takes on more of the work, the scarce skills are setting clear direction, defining intent and applying judgment. People and agents work side by side. AI handles more of the execution, and people decide what is worth building and why.
The goal is for every team to do more with the same resources, move faster and work better together, without compromising quality.
You’ve argued that AI-driven productivity gains can quickly disappear when the underlying workflows are inefficient. What are the most common operational bottlenecks that prevent companies from realizing the full benefits of AI, and how should leaders approach redesigning these workflows?
Speeding up individual tasks doesn’t automatically speed up the organization. If processes, dependencies and communication between teams stay the same, most of the productivity AI creates gets lost along the way.
The most common bottlenecks aren’t technical. They’re unclear requirements that cause repeated rework, slow handoffs between teams and work that sits waiting for approvals. AI makes each step faster, but the time is lost between the steps.
Leaders should start by measuring where time actually goes, because the real bottleneck is often not the one everyone assumes. Then redesign the workflow from start to finish, fixing the handoffs, definitions and feedback loops rather than only individual tasks. Finally, connect every AI effort to the KPIs teams already own.
Most importantly, leaders need to change the mindset. Automation and acceleration should become the default. Instead of accepting existing processes, teams should keep challenging how things are done so they can move faster and increase their impact without compromising quality.
One of the principles you emphasize is the importance of clearly defined requirements before introducing AI agents into a workflow. How should organizations rethink the way they define tasks, communicate expectations, and establish success criteria when AI agents are responsible for executing significant portions of the work?
When AI agents do a large part of the work, organizations need to rethink how they communicate and define tasks. A vague request doesn’t stop an agent. It guesses, and every wrong guess leads to more rework.
I think about it in three layers:
Clear tasks, goals and context. The agent needs to understand what needs to be done, why it matters and the context around it. Before the work starts, it should be able to ask about anything unclear, so gaps are resolved upfront instead of discovered halfway through.
Self-validation. Success criteria should be defined in advance, in a way the agent can check on its own. In software, for example, an agent should be able to run the code, test that it works as intended and confirm that it passes the tests before anyone reviews it.
Humans as the final layer. People should receive a result the agent has already refined over several iterations against the agreed criteria. Their role is to apply judgment, not to fix basic mistakes.
The better the definition, the less human time each result needs.
You advocate enabling AI agents to validate their own work before it reaches human reviewers. What does an effective AI-driven validation process look like in practice, and how can organizations ensure that automated quality checks don’t simply reinforce the same mistakes made during execution?
An effective validation process tests the work the way it will actually be used. The agent should run the real system from start to finish, confirm the result matches the success criteria and fix what fails before anything reaches a person.
To keep validation from repeating the agent’s own mistakes, a few principles matter:
Separate the criteria from the execution. People should define what success looks like before the work starts. If an agent decides what “done” means, it will grade its own homework.
Set hard guardrails. Not every check should depend on an agent’s judgment. Some must be fixed, deterministic tests that verify specific parameters, standards and policies. These give the agent clear boundaries it can’t reason its way around.
Use an independent reviewer. A second agent, with fresh context and no stake in the original approach, catches errors the first one misses.
Show evidence, not just a verdict. Test results, screenshots and recordings let people see what happened instead of trusting a “pass.”
Human attention then goes where the risk is highest. The goal isn’t to remove people from the process. It’s to make sure their time goes to judgment, not mechanical checks.
Traditional organizations often rely on multiple layers of management, approvals, and handoffs between teams. How does an AI-native operating model change the way engineering, product, security, and business teams collaborate, and what happens to traditional roles and responsibilities as agents take on more of the execution?
AI lets every person and every team produce more. But if handoffs, approvals and coordination stay the same, that extra output just piles up waiting for the next step. The organization doesn’t move faster. It creates new bottlenecks. An AI-native operating model has to redesign how teams work together, not only how each team works.
That starts with how information moves. More of it should flow through automation, so meetings are saved for decisions rather than updates. When insights from the field reach product directly, decisions become faster and more accurate, and that carries through to clearer definitions and better execution.
Approvals need the same rethinking, and this is the hardest part, because it means changing what the organization treats as normal. Approvals still matter, but they shouldn’t treat every change the same way. For example, adding a risk assessment to a specific approval process can reduce manual approvals for low-risk work while keeping human review where it counts. That’s just one way to speed up a process without giving up control.
Roles are being redefined as well. As agents take on more of the execution, people move from owning tasks to owning outcomes. They need to understand the full context, challenge assumptions and step beyond their traditional boundaries. The organizations that win will be the ones that redesign how people work together, not only the ones that add the most AI.
Ocean itself uses specialized AI agents to investigate sophisticated email threats. How has your experience developing agentic AI for cybersecurity influenced the way you deploy AI internally, and are there lessons from building autonomous security systems that apply to managing an AI-native organization?
Building autonomous agents for email security shaped how we use AI internally in two ways.
The first lesson is that agents perform only as well as their understanding of context. A security agent makes better decisions when it understands the organization it protects. The same is true internally. To get good results from agents, we need to give them a real understanding of how the organization works, what each task is for and how it connects to the bigger picture.
The second is that agents need continuous evaluation. In production, we constantly evaluate how our agents perform, where they fail and which tools or data they’re missing. Internally, we apply the same approach. We need a dedicated place to test, train and get feedback on agents. Without it, it’s hard to tell whether an agent is improving or what it needs next.
Running an AI-native organization is a lot like building an autonomous security system: give agents deep context, keep evaluating how they perform, and keep people in the loop where judgment counts.
As organizations give AI agents greater autonomy, concerns arise around accountability, security, and unintended consequences. Drawing on your cybersecurity background, how should companies establish guardrails that allow agents to operate independently without compromising security or losing meaningful human oversight?
Autonomy is where agents are heading, and organizations should get there step by step. Start where the impact is well understood, prove the value and then expand.
Real guardrails aren’t instructions. Telling an agent “don’t do this” isn’t a security control. Guardrails need to be enforced: strict limits on what the agent can access, verification that it can’t reach anything outside its scope, and continuous monitoring that blocks actions nobody planned for. Like any other part of the system, they should be tested and validated regularly.
Accountability starts with the operator. The person running an agent defines its boundaries: what it can do on its own, what needs approval and where it can’t go. Take an agent that responds to email threats. It can quarantine suspicious messages on its own, quickly and reliably. For actions with wider impact, such as blocking a sender across the entire organization, the operator may start with a review step. As the agent’s track record grows, those boundaries can expand with confidence. Clear guardrails are what make greater autonomy possible.
This matters even more because attackers are increasingly targeting AI agents. The business wants to move faster, but security teams still don’t have a clear playbook for protecting agents. That’s why the foundation has to be enforced permissions and constant monitoring.
Many companies measure AI adoption through time saved, tasks automated, or employee productivity. What metrics do you believe actually demonstrate whether an AI-native operating model is delivering sustainable business value, particularly when accounting for rework, quality, and coordination costs?
Time saved is an important metric, and it’s often the first clear sign that AI is creating value. But it shows its full impact when it’s paired with metrics that show what the organization does with that time, and whether the gains hold up over the long term.
Quality. Faster is only better if quality holds or improves. Track rework, defects and how often results need to be corrected after delivery. When speed and quality improve together, the gains are real and lasting.
Business expansion. The real value of an AI-native organization is what the time it saves makes possible. We measure how many new capabilities and initiatives we take on without adding headcount.
End-to-end speed. How long an agent takes to develop a feature is one part of the picture. The real measure is the full process, including every handoff, from the moment we decide to prioritize something to the moment it’s live in the product. That’s how fast we actually run, and it shows the coordination costs that task-level metrics miss. In security, end-to-end speed is critical. Attackers are already using AI agents, so the time it takes to detect and respond to a threat has to keep shrinking.
Time saved shows that AI is working. These metrics show it’s building lasting value for the business.
You’ve suggested that smaller, more agile technology companies can gain a meaningful competitive advantage by redesigning their operations around AI. What advantages do AI-native startups have over established incumbents, and how do you envision the structure of successful technology companies evolving over the next three to five years?
AI-native startups have a structural advantage. They can build workflows, technology and teams around AI agents from day one, while incumbents must deal with decades of legacy infrastructure and organizational complexity. The bigger gap isn’t technical. Everyone has access to the same AI tools. The advantage belongs to companies that can redesign how they work, without layers of approvals, handoffs and fixed roles to unwind first.
Over the next three to five years, I believe successful technology companies will move beyond basic AI automation toward agentic operating models, where specialized AI agents execute entire workflows. Human teams won’t disappear, but people will spend less time executing tasks by hand and more time providing context, setting guardrails and directing outcomes. The result will be leaner organizations that can scale faster, operate more efficiently and compete with companies many times their size.
Thank you for the great interview, readers who wish to learn more about agentic email security should visit Ocean Security.












