The Security Stack Is Becoming an LLM. Insider Risk Still Isn’t.

The other week’s market reaction said the quiet part out loud: the future security stack is going to be run by LLMs, whether that’s Claude, Gemini, OpenAI, or whichever model or combination of models wins the next cycle. When Anthropic debuted Claude Code Security, a tool positioned to autonomously scan codebases for vulnerabilities and suggest patches, cybersecurity stocks sold off hard. Investors didn’t just hear “new feature.” They heard “new operating layer.” 

Many of us saw this shift coming. AI won’t simply augment point tools. It will absorb them. The market’s worries are a preview of a structural change. Detection and response workflows that used to be spread across dashboards, rules engines, and analyst queues will increasingly collapse into model-driven reasoning and execution. Anthropic itself frames this as a near-term inevitability, expecting a large amount of the world’s code to be scanned by AI.  

But here’s the part the market narrative misses. LLM-as-security-stack does not equal LLM-solves-security. Not fully. Not safely. And definitely not for insider risk. 

Yes, the LLM will run the stack.  

If you squint, Claude Code Security is the natural endpoint of a trend. LLMs moving from chat to agents, from “tell me what’s wrong” to “find it, validate it, fix it.” Anthropic says Claude Code Security is designed to catch subtle logic flaws and route findings through verification to reduce false positives.  

That’s powerful. It’s also why investors immediately started asking whether legacy suites become unnecessary. The market doesn’t panic because it’s irrational. It panics because it recognizes things have changed. 

This is exactly what we’ve been talking about in the Shadow AI conversation. AI adoption isn’t waiting for governance. It’s moving at copy/paste speed. Agent speed. In my Shadow AI Is Everywhere post, I highlighted Gartner® survey findings that 69% of security leaders suspect or have evidence employees use prohibited public GenAI at work, and Gartner projects that by 2030, over 40% of enterprises will face incidents tied to unauthorized GenAI use. 

When the stack becomes an LLM, security becomes an interaction model. Humans’ prompting, agents acting, and data moving through systems that traditional controls weren’t built to observe.  

But insider threats are not a reasoning problem. 

Here’s the critical distinction. LLMs are great at reasoning over what you give them. Anthropic can build a model that finds bugs. But insider risk is a different class of problems. It’s not “identify vulnerability X in code Y.” Insider risk depends on seeing intent, not just activity, and meaning, not just signals.  

Effective IRM is based on high-fidelity, privacy-safe, longitudinal signals about human activity across time, tools, and context. It’s:

• What does normal behavior look like for this person/team/role, not on one day but over the course of his/her time at the organization? 

• What changed, and why? And can I go beyond the activity/alert to see meaning, not just signals? 

• Do I know if these behavior changes are benign, negligent, coerced, or malicious? How do I know? 

• When tasks are automated by AI agents, what does normal agent behavior look like over time, and how does it differ from human‑initiated activity? 

• What action reduces risk without breaking productivity? 

DTEX CEO Marshall Heilman has been clear that insider risk requires behavioral intelligence, early indicators, and proactive detection—not simply reactive controls at the point of exfiltration.  

That’s exactly why DTEX’s differentiation is the data we collect plus intent mapping at scale. The hard part isn’t writing a smart summary after the fact. The hard part is collecting the right signals continuously to identify risk early, without turning security into surveillance.   

High-fidelity telemetry (privacy-safe) is the moat.

DTEX is built around the industry’s most reliable and comprehensive behavioral intelligence, capturing rich activity metadata, maintaining privacy protections, and correlating across actions to identify risk early. And this is not “hand-wavy”. DTEX collects metadata whether it is interesting at the time or not, paired with contextual enrichment and risk modeling to correlate a range of activities into a single risk score.  

That matters in the AI era because the organization’s challenge isn’t “do we have the right LLM?”. LLMs are only as good as the data they can access. The challenge really is  
Do we have [enough] trustworthy, governable, bias-resistant data about what’s happening in the environment, human and non-human? 

Without good data, even the smartest model in the world is still guessing, forced to make assumptions and increasing the risk of wrong conclusions.  

Agentic AI introduces a new insider class: non-human actors with broad access.  

The most important downstream consequence implication of Claude Code Security isn’t code scanning. It’s what it represents: autonomous tool use embedded in enterprise workflows. 

The DTEX i³ research has been tracking the rise and risks of AI agents, including the core enterprise problem of distinguishing human-initiated actions from agent-initiated actions, which complicates auditing, accountability, and compliance.  

This is where DTEX sees the next frontier: independently risk scoring agentic AI beyond traditional guardrails. Guardrails are necessary, but they’re not sufficient when: 

• agents operate across apps and identities, 

• actions look “authorized” because permissions are valid, and 

• speed collapses the time window for intervention. 

Behavior is still the signal. Only now, you’re tracking behavior across humans + agents.  

Why can’t Anthropic just solve insider threat without DTEX? 

Even if Claude (or any model) becomes the UI for the entire security stack, insider risk still depends on foundational capabilities DTEX has invested in for years: 

1. Longitudinal behavioral baselines (not just event streams)  

2. Privacy-by-design data handling (pseudonymization, bias reduction) 

3. Risk-adaptive scoring and response that unifies human, data, and AI risk into actionable prioritization  

4. GenAI oversight where work happens including visibility into what employees upload to AI services and mechanisms to prevent sensitive data leakage  

DTEX has already positioned this as a core insider risk requirement, understanding what employees upload to GenAI platforms and mitigating IP exposure while maintaining privacy standards.  

So yes, LLMs will run the stack. But DTEX supplies the ground truth, which means the behavioral telemetry and intent-layer understanding that lets an AI-driven stack make the right call. 

The takeaway: AI-first security needs a control plane. DTEX is that plane. 

The market saw Anthropic’s announcement and asked, “Will AI replace cybersecurity vendors?” DTEX’s answer is more precise: 

• AI will replace parts of the stack that are primarily reasoning + automation.  

• Insider risk will increasingly be the differentiator because it’s fundamentally behavioral, contextual, and longitudinal.  

• Agentic AI expands the insider problem to non-human actors, making independent risk scoring essential.  

Or how about this? Think of the AI-driven security stack as a police investigation. The LLM is the investigator responsible for reasoning, asking questions, and forming hypotheses. DTEX data is the case file, providing situational context across time, tools, and actors. DTEX Intelligence is the investigative tradecraft: the domain knowledge that enables the investigator to interpret evidence correctly and reach defensible conclusions.

If you want to safely adopt AI agents, govern Shadow AI, and keep humans productive while protecting your crown jewels, you need visibility into what’s happening and the ability to map it to intent before it becomes an incident.  Reach out to us for a demo and let us show you.