Every legacy security vendor seems to be slapping semantic filters onto their pre-existing products and promoting them as actual defense. A rudimentary classifier here or there, or a quick embedding lookup, an MCP inspection, and suddenly you have semantic security ready to pitch to an InfoSec team that is quickly becoming skeptical. Anyone who has spent time in the trenches knows how weak that is. And hackers are well aware of how simple linguistic misdirection, wordplay, and asynchronous payloads slice right through those barriers.
The security swirl inside the vendor community needs to move past a consumer chatbot mindset and confront real enterprise security if they want any hope of containing the risks. Semantics alone will never get you there, and I talked about this industry skill gap recently in an article about Cavemen with Blowtorches. It triggered a bit of backlash but a lot of healthy debate and the gap remains as wide as ever.
Once you step away from the chatbot hellscape and get down to brass tacks, the enterprise requires industrial-strength protection that most vendors are not providing. What we are seeing today is just another cycle of AI washing, selling quick fixes, the bandaid, to complex architectural problems. Tacking semantics onto a firewall or an API gateway is a feel-good exercise no different than attempting to bake safety into model weights. None of it protects the enterprise. None of it keeps the crown jewels safe.
Don’t get me wrong, all those traditional security measures and capabilities are fantastic and necessary, but AI is a new threat, and we see that every day in the news.
Enterprises need hard runtime control, not polite guardrails.
It’s all Semantics
The irony is that a critic might suspect I’m just playing with semantics myself when writing this article. But the foundational threat with large language models is not semantics. What gets sold as semantics in most security products is little more than syntax and vocabulary dressed up with statistical inference, the raw building blocks for linguistic games.
I realize savvy readers will push back on that characterization, and technically they should. But the implementation reality inside most vendor solutions is semantic filtering. A classifier. Some intent detection. A probabilistic scorer. It is nowhere close to semantic interpretation, never mind the real threat.
The real operational threat is that AI models are brilliant at semantic association. Association is the word that actually matters in this statement. It represents both the core capability of these systems and their primary vulnerability. It is not about the words themselves. It is about how the engine connects them and AI is brilliant at that.
Context is another one of those terms the industry throws around without thinking. And one I’ve written about recently in Asleep at the Wheel where AI is terrible at context. Dumping raw context into a prompt window comes nowhere close to Context Competence, where competence is the most important part of this equation.
Now put association and competence together and a hacker has an absolute gold mine in which to launch an attack. They can quickly construct an automated engine capable of siphoning details, bridging separated silos, and exposing every architectural weakness across your enterprise infrastructure. Traditional security vendors are not even tracking association or competence, which means they are completely blind to the actual attack surface of AI.
Hidden Dangers of Shadow AI
Then there is the innocent trap sitting right inside the building.
Because AI is brilliant at association and your internal workforce has immense competence in running the business. The combination becomes deadly when employees build agents on top of Copilot, Clawdbot or other internal tooling to automate their daily work, that competence collides directly with the very risks facing an organization.
The outcome is quiet, completely innocent exfiltration … the data loss that has accelerated at machine speed over the last year. There’s no malicious intent on the part of the employee or even the agent they build. Prompts are clean and innocent looking and the actions seem in line with the productivity boost that everyone wants. Including myself.
But, the agent is devoid of risk, of compliance, of policy awareness, of consequences, of tribal knowledge that is ultimately stuck with the human. And the agent, with its brilliant ability to make associations, quietly aggregates the sensitive data available in documents, emails and systems, and acts innocently to send it on its merry way. Sensitive enterprise and customer data is now out in the ether … the very data that the organization spent years protecting.
Unlike consumer chatbots, enterprise security has to handle the reality of sophisticated multi-threaded, multi-channel, asynchronous, and asymmetric attacks. More than that, however, it has to handle operational weaknesses that are also multi-threaded, multi-channel, asynchronous, and asymmetric. Not every breach or exfiltration stems from a hostile intruder. A huge portion comes from well-intentioned humans and autonomous agents simply trying to get their jobs done.
The Control Plane Necessity in Security
A genuine control plane requires embedded intelligence designed to govern association and competence directly. It has to exist at the edge and deep within the bones of the infrastructure to contain real threats. That means enforcing runtime control as data continuously aggregates, processes, transforms, and moves across the enterprise.
-
Deeply integrate control into orchestration frameworks and execution flows rather than treating security as an external proxy.
-
Protect against contextual drift, state mutations, and multi-channel vectors across the entire lifecycle of an interaction.
-
Maintain verified lineage tracking for all data, following it from at-rest storage through every in-motion transformation.
-
Enforce fine-grained policy at the transactional judgment level instead of stopping at basic identity and authorization checks.
-
Apply strict containment boundaries to agent transactions and data flows to prevent unauthorized state changes and policy breaches.
-
Build context competence awareness into the system to identify sensitive aggregations, triggering automatic redaction, generalization, or lockdown before critical data moves across trust boundaries.
Semantic inspection is not AI security, and inspecting MCP traffic is too little, too late … you might as well classify the data as already lost.
Protecting the enterprise crown jewels takes far more than semantic hand-waving. It demands cradle-to-grave runtime control where policy enforcement travels everywhere the data goes.

