COMMENTARY: The AI experimentation phase is over. Organizations spent 2024 and 2025 deploying copilots, testing agents, and letting employees try every AI tool that promised productivity gains.Some delivered. Many didn't.Now comes the reckoning, where security teams separate the tools that actually work in production from the ones that just sounded good in demos.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts.Read more Perspectives here.]
In cybersecurity specifically, we're watching this play out in real time. The vendors who promised autonomous AI would eliminate alert fatigue and run SOC operations overnight created expectations they couldn't meet. That failure doesn't mean AI can't help. It means we learned what actually works. And what works isn't a general-purpose autopilot — it's purpose-built agents designed for specific, measurable tasks within defined workflows.
From experimentation to operational deployment
The defining shift in 2026 is operational deployment. Not proof-of-concept. Not pilot programs. Actual integration into production workflows with measurable outcomes and clear success metrics. This requires a fundamentally different approach than what dominated the last two years.We're seeing the industry split between vendors still selling autonomous fantasies and practitioners building focused agents that work. The difference? General-purpose AI tools promised to do everything but struggled to do anything particularly well. Purpose-built agents are designed for specific domains (phishing detection, incident simulation, SOC triage) where they can be trained on relevant data, measured on clear outcomes, and improved through focused iteration.
This isn't a retreat from AI. It's a maturation. Security teams don't need an AI that can write poetry and also maybe detect threats. They need agents that excel at the specific tasks that consume their time. Analyzing suspicious emails. Simulating attack scenarios. Threat hunting.When you design for a focused outcome, you can actually achieve it.
Human expertise doesn't compete with AI, it validates it
There's growing skepticism around AI-generated content, and for good reason. When everything from threat intelligence to security blog posts gets auto-generated, people tune out. The signal-to-noise ratio collapses. Security teams don't trust analysis that might have been hallucinated by an LLM.The solution isn't to abandon AI. It's to make human validation visible and verifiable. The most effective AI implementations aren't autonomous systems operating in isolation. They're human-in-the-loop models where analysts validate what AI flags, provide feedback that improves the models, and serve as the quality assurance layer that keeps automated systems trustworthy.This becomes a competitive differentiator in 2026. Products that can prove real human expertise is in the loop (i.e., reviewing AI decisions, training on actual threats, providing oversight) will stand out. We learned that AI plus human judgment outperforms either one alone.
AI's impact on the threat landscape can't be overstated
I've been tracking AI threat intelligence reports closely, and what I'm seeing represents a fundamental shift in who can be a cybercriminal. When I analyzed Anthropic's threat intelligence findings earlier this year, the pattern that emerged goes way beyond attack sophistication. AI is creating threat actors who shouldn't exist.We're past the point where AI augments existing criminal capabilities. We're now seeing AI create criminal capabilities that didn't exist before, operated by people who couldn't have been criminals in the traditional sense.The North Korean IT workers maintaining Fortune 500 engineering positions while needing AI assistance for basic technical tasks. The ransomware operator selling $400 to $1,200 malware packages despite being completely unable to implement encryption algorithms without AI guidance. These aren't skilled attackers using AI to work faster. They're people with ZERO baseline capability becoming sophisticated threat actors purely through AI dependency.This demolishes traditional assumptions about cybersecurity defense. Our entire defensive model has always assumed some baseline of human capability and scaled protective measures accordingly. When AI provides unlimited capability to actors with minimal skills, that model breaks. Technical skill is no longer a prerequisite for sophisticated attacks, and the barrier to entry for cybercrime operations has essentially vanished.The numbers from a survey from Osterman Research bear this out. Eighty-eight percent of organizations experienced at least one AI-enabled security incident that undermined trust in digital communications within the past 12 months. At the same time, 60% reported having low confidence in their organization's ability to defend against such attacks. That confidence gap isn't just about attack sophistication. It's about defending against an entirely new category of threat actor that traditional security controls weren't designed to stop.
Blind spot I’m worried about
Everything I've seen documented so far involves hosted AI services where providers could eventually detect and ban malicious usage. But as capable AI models become available for self-hosting (with no safety guardrails and no provider oversight), that last line of defense disappears entirely. Yeah, I’m talking about you, DeepSeek, and the risks you bring to the party. The dependency-driven criminals operating today through monitored services could operate tomorrow with complete impunity using downloaded models that have zero safety and detection mechanisms.
Focused agents and measurable outcomes (this is the way)
The reset in expectations isn't a failure. It's a necessary correction that clears the way for AI that actually works. Organizations need to shift from asking "Can AI solve all our security problems?" to "Which specific security processes can purpose-built AI agents measurably improve?"That's a different question with different answers. Instead of one autonomous system trying to do everything, you deploy focused agents. One that specializes in identifying sophisticated phishing attempts. Another that simulates attack scenarios to test defenses. Another that triages SOC alerts based on actual risk. Each agent operates in a defined domain where it can be trained on relevant data, measured on clear metrics, and continuously improved based on feedback from security analysts.This approach acknowledges what the hype cycle obscured. AI is a tool, not a silver bullet. Used correctly (deployed in focused applications with human oversight), it delivers significant value. Used incorrectly (oversold as a general-purpose replacement for human judgment), it creates disappointment and skepticism.For the sake of everyone in cybersecurity, we need to be honest about what AI can do. Not what it might do someday. Not what makes for impressive demos. What it can demonstrably accomplish in production environments today. That clarity will separate the vendors building real solutions from those still selling futures that may never arrive. And for organizations defending against increasingly sophisticated AI-powered attacks, that distinction matters more than ever.
Audian Paxson is principal technical strategist at IRONSCALES.Paxson is a recognized authority in enterprise IT infrastructure and cybersecurity with over 20 years of experience driving innovation in cloud security and advanced threat protection. He holds three USPTO patents focused on groundbreaking advancements in enterprise security and is known for his expertise in leveraging AI to counter emerging threats. Paxson’s deep understanding of the rapidly evolving threat landscape and his ability to bridge technical insights with practical applications make him a sought-after voice in the cybersecurity industry.
