For over a decade, I worked in application security. Helping companies defend against DDoS, application-layer vulnerabilities, and automated attacks: credential stuffing, web scraping, enumeration, fake account creation. That world shaped how I think about attackers, not as distant, abstract threats but as adversarial teams with their own profit motivations and iteration cycles, their own tooling, and their own version of a deployment pipeline.
What I observed over the last few years in that space was genuinely staggering. And a lot of it came down to one challenge: bot detection logic must live in the browser to be effective. In the open. For anyone to inspect and attempt to decipher. You needed time, curiosity, and enough patience to study what defenders had built. A collaborative underground community made sure that work got shared. An adversarial cat and mouse game that the companies I worked for played better than anyone else.
Then it got worse.
Once capable LLMs became widely available, the remaining barriers to effective attacks started collapsing. Attackers could build stealth bots without writing code. They could feed obfuscated JavaScript into a model to help understand what it was doing. Skills that previously took months to develop could be approximated in an afternoon.
What made this genuinely alarming wasn’t just the sophistication. It was what it did to the tempo. The attack cycle kept compressing: attempt, analyze, adjust, retry. Faster and faster. While defenders were operating on human timelines and the cognitive limits of working hours, attackers were iterating almost continuously, and sharing their findings across the community as they went. The asymmetry wasn’t new. AI made it accelerate.
The answer on the defense side moved towards polymorphism to disrupt the attacker’s iteration cycle, AI as a critical layer for detection, and automating the enforcement security teams would otherwise have to apply manually. You can’t stop a continuously iterating attacker armed with AI without AI. The defense has to have its own loop.
That’s the context I carried into every conversation I had about what to do next.
The signal
When I first heard about the opportunity at Bricklayer, the immediate reaction was recognition. They were building an agentic cybersecurity platform where AI agents work together to help defenders move faster across security operations. The near-term focus was security operations: alert triage, investigation, response, vulnerability management, threat intelligence, and threat hunting. But the architecture wasn’t limited to that. There was no technical reason the same platform couldn’t extend into other areas of cybersecurity, IT, or beyond. That kind of extensibility matters a lot when you’re thinking about where the problem is heading.
The Inflection Point
A few months into the role, something happened that confirmed everything I’d been watching build.
Early testing of models like Mythos has shown a meaningful uplift to identify real vulnerabilities. And frontier models like GPT-5.5 can now break down complex problems, use tools, and work through multi-step investigations toward an outcome. The trajectory is clear: these models will lower the cost of exploitation and increase the volume and complexity of what security teams have to respond to. And now those same models are powering agentic attacks that can operate autonomously and continuously. The kind of relentless, iterative pressure I watched play out in bot defense as a result of AI is now the reality across cybersecurity at large. The tipping point isn’t coming. It’s here.
Continuous, AI-assisted attacks against organizations aren’t a future scenario. Which means defenders need to operate continuously too. And right now, most can’t.
The staffing math doesn’t support it. The tooling wasn’t built for it. Static workflows, by definition, can’t adapt to it.
The Hard Part
What I’ve learned at Bricklayer in a short time is that the hard problem isn’t building agents that are faster. There’s no shortage of those. It’s building a system where agents can coordinate without creating more noise than they resolve. That’s a different engineering challenge entirely, and it’s where most agentic AI proposals in security fall apart.
If you deploy agents without shared context, they duplicate work and miss connections. If you deploy them without coordination, you get parallel investigations that step on each other. If you deploy them without governance, you lose the audit trail and policy enforcement that security operations require by definition. Speed without structure doesn’t improve mean time to respond. It makes the chaos harder to manage.
The real value in a well-built agentic AI platform isn’t any single agent. It’s the orchestration: structured context flowing across agents, coordination so work progresses logically, guardrails that enforce policy without requiring a human to approve every step. That’s what compresses dwell time. That’s what allows MTTR improvements that actually show up in outcomes. That’s what drives consistent, thorough investigations that reduce risk exposure over time.
But it only works if it spans the full spectrum of security operations rather than solving one piece of the workflow in isolation, because those pieces are increasingly intertwined. The industry is guilty of creating AI silos, point solutions that do one thing well but don’t talk to each other and require analysts to do the integration work manually. Agentic AI implemented the same way just gives you smarter silos. It doesn’t change the system.
Closing the Loop on Cyber Defense
Agentic security operations is one of the most invested areas in cybersecurity right now, and for good reason. It’s not a trend. It’s a structural response to a structural problem: the attack surface is expanding faster than human capacity to monitor and respond to it, and the pace of attacks and speed at which new exploits are discovered is compressing the time windows defenders have to act.
Every area of cybersecurity is going to need agents. The question isn’t whether the shift happens. The question is whether the organizations building these systems get the hard parts right: the coordination, the governance, the trust-building that has to happen before security teams will let autonomous systems take meaningful action.
That last part matters more than it gets credit for. Full autonomy in cyber defense isn’t where this starts. It’s where it might eventually get to, after the technology has earned the right to operate there. The foreseeable architecture, controlled autonomy with human oversight, coordinated agents that surface findings and recommend actions rather than take them unilaterally, is what makes agentic security deployable in real enterprise environments today. Working with the tools organizations already use, and strengthening the security talent they already have.
The shift to agentic defense isn’t a criticism of what came before. The tools built over the last decade were genuinely good, and the people building them were brilliant. But the threat has crossed a threshold where the entire operating model for cybersecurity defense has to evolve. That’s the problem I wanted to work on next.
That’s what Bricklayer is building. And that’s why I joined.
