"Agentic" Was Everywhere at RSAC 2026. The Real Story Was the Exhaustion.

The Lemonade Problem Nobody Is Talking About

Doug White introduced a framing during the Day 3 broadcast that stopped me — he called it the "market for lemonade" [1]. Here's how it works: You buy a security product that generates massive amounts of data (vulnerability scanners, SIEMs, threat intel feeds). Because you're drowning in that data, you buy another product — often now powered by AI — to help you make sense of it. But nobody stops to ask whether the original data was any good. Maybe your scanner is only covering 50% of your assets. Maybe your alerts are misconfigured. You hand that garbage to an AI that confidently produces insights, recommendations, and risk scores. You've made lemonade out of bad lemons. The AI didn't fix the problem. It just dressed it up. This is the fundamental challenge for AI in the SOC right now. The vendors selling "AI for security operations" are often bolting sophisticated models onto organizations that haven't solved basic data hygiene. And when the AI tells you there are a million critical vulnerabilities on 30 scanned hosts, that's not an AI problem — that's a measurement problem the AI just made more embarrassing [1].

"Agentic" Means Nobody Knows Who's in Your Network Anymore

Here's the part of the RSAC conversation that I think deserves more coverage from a developer perspective. One of the security leaders Sinabria interviewed on Day 2 floated an idea that stuck: in a world full of agentic AI — ephemeral bots that spin up, do a task, and disappear — traditional identity management breaks down completely [1]. Right now, your IT team knows roughly who has access to what. A human employee builds up a history. Trust is earned over time. But agentic AI creates what one security professional at the conference called "a Kubernetes of personalities" — hundreds of ephemeral identities that appear when they're needed, request access to systems, do work, and vanish. How do you authenticate an agent that doesn't exist long enough to build a history? The emerging answer, according to multiple RSAC speakers, is reputation over identity: rather than asking "who is this?", you ask "does this type of agent, in this context, from this source, have a track record of behaving safely?" [1] That's a solvable problem in theory. In practice, nobody has built the infrastructure for it at scale yet. And it's coming fast — right now, every major AI platform (Claude, Copilot, Gemini) is asking for broader and broader system access to be more helpful. Security teams are already getting calls from executives who've quietly installed AI tools on dozens of systems and are now asking for retroactive approval.

You'll have an army of agentic AI to deal with the army of agentic AI that's facing you.

— Security leader at RSAC 2026, as recounted by Adrien Sinabria, Cyber Risk TV

2026 Is the "Put Up or Shut Up" Year

The 451 Research breakfast at RSAC apparently had a financial analyst in the room — someone who looks at cybersecurity purely as a market category, not as a technical discipline. His read: last year was the biggest M&A year in security history (the Wiz acquisition alone was larger than entire previous years of deal activity). This year? Down. A lot of the AI security startups that didn't get acquired in 2025 are going to struggle to raise a Series B. And the reason isn't just a slow market — it's that the "AI for SOC" story hasn't converted from pilot programs to production deployments fast enough [1]. Sinabria put it plainly: organizations are running out of patience with experimental AI budgets. You've been spending on AI security tools for two or three years. You've run proofs of concept. You've had vendors tell you "any day now." And the board and the CFO are starting to ask a straightforward question: where's the measurable value? If you can't answer that, the budget conversation in Q3 2026 is going to be brutal for a lot of AI security vendors [1]. For builders and developers working on security tooling, this is actually good news in the long run. The market is forcing evidence over promise. Vendors who can show a clean, auditable reduction in mean time to detect or mean time to respond — with real numbers, not testimonials — will survive. Everyone else is going to get rationalized out of the stack.

The Tech Debt Bomb Nobody Wants to Defuse

One of the sharpest observations from the Day 3 broadcast: regardless of how the AI-in-security bet plays out, the industry is about to speedrun ten years of tech debt in one year [1]. Think about what that means. If AI wins and organizations fully commit to AI-driven SOC automation, they'll have a decade's worth of governance gaps, integration debt, and untested edge cases compressed into 12 months. If AI loses — if the tools don't deliver and organizations pull back — they'll have a decade's worth of ripped-out implementations and mismatched workflows to clean up. There's no clean exit from the AI experimentation phase. The debt is coming either way. The only variable is whether you're managing it proactively or you're the next generation's mop-up project.

What's Actually Working

Strip away the agentic hype and here's what the RSAC conversations actually pointed to as areas with real traction: Threat detection at scale: AI models processing log volume that human analysts can't handle are genuinely useful, especially for finding low-and-slow attack patterns that drown in alert noise. AI as first-pass triage: The framing that came up repeatedly is "AI as a great assistant" — it goes and grabs research, compiles context, flags anomalies — but humans stay in the decision loop for high-stakes calls [1]. Identity governance for non-human entities: The agentic identity problem is real, but it's also a solvable engineering problem. Vendors who get ahead of it — building reputation-based trust frameworks for AI agents — have a legitimate market to go after. What's not working, at least not yet: fully autonomous AI making security decisions without human oversight in production environments. The concerns aren't philosophical — they're practical. Data quality isn't good enough. Explainability isn't good enough. The institutional trust isn't there. The organizations that are handing AI the wheel and walking away are the ones who'll have the most interesting post-mortems [1].

The people at RSAC 2026 who know this space most deeply — not the vendors, but the practitioners who run actual SOCs, who've been doing this through every hype cycle since the early 2000s — sound like experienced engineers at the start of a hard project. They know something big is changing. They're not sure exactly what their job looks like on the other side of the transition. But they're not panicking, and they're not drinking the Kool-Aid. They're testing carefully, measuring everything, and waiting for the tools to prove themselves. That's not skepticism — that's professionalism. And honestly? It's the right call [1][2].