AI Companies Build Models. Cybersecurity Companies Defend Systems.
AI companies build models.
Cybersecurity companies defend systems.
Those are different jobs. Treating them as interchangeable is a category error that organisations keep making, usually under pressure to “do something about AI risk” without examining who is actually equipped to do it.
An AI company optimises for capability, performance, product velocity and competitive positioning. Its engineers are trained to improve accuracy, reduce latency, expand context windows, and ship features. When security appears in their work, it is usually as a constraint on the product roadmap or as a compliance checkbox attached to a release.
A cybersecurity company optimises for residual risk, adversarial resilience, detection, response, and the long-term cost of being wrong. Its practitioners are trained to assume the system will be attacked, to map attack surfaces that the builders did not intend, and to design controls that survive contact with an opponent who has time, incentive and creativity.
These are not the same skill set. They are not even the same professional culture.
The problem is not the model. It is the system around the model.
Most of the serious security failures involving AI systems do not originate in the core model weights. They originate in how the model is integrated, how it is given tools and privileges, how its outputs are trusted, how its training and fine-tuning data are sourced, how its agents are allowed to act, and how the organisation monitors and contains the resulting behaviour.
Prompt injection, excessive agency, tool over-permissioning, insecure retrieval pipelines, poisoned training or fine-tuning data, model extraction, and supply-chain compromise of the model-serving stack are security problems. They are solved by threat modelling, least privilege, input/output validation, monitoring for anomalous behaviour, secure development practices, and incident response capability.
An AI company can improve the model’s robustness on certain benchmarks. It cannot replace the defensive engineering required to run that model safely inside a real organisation that has real assets and real adversaries.
Different incentives produce different outcomes
An AI vendor’s primary commercial incentive is to demonstrate capability and close deals. Security claims are marketing assets. When a vulnerability is found in their product, the institutional response is shaped by product reputation, release schedules, and competitive pressure.
A cybersecurity firm’s primary commercial incentive is to reduce the client’s residual risk and to remain credible when things go wrong. Its reputation depends on accurate assessment, not on optimistic capability narratives. That difference in incentive structure is not subtle. It determines what gets prioritised, what gets deferred, and what gets honestly reported.
Hiring the builder to audit the building is rarely the strongest arrangement.
What “fixing AI cybersecurity issues” actually requires
It requires the ability to:
Map the full attack surface of the deployed system, including the model, the orchestration layer, the data pipelines, the tools the model can call, and the humans who trust its outputs.
Distinguish laboratory robustness claims from operational security posture.
Design controls that degrade gracefully under attack rather than controls that only work when the attacker cooperates with the threat model.
Integrate detection and response for novel failure modes that traditional SIEM and EDR were not built to see.
Advise on liability, insurance, regulatory exposure and residual risk in language that boards and legal teams can act on.
These are cybersecurity competencies. They are not automatic side-effects of having trained a large language model.
A concrete local example
In South Africa, firms such as Cybersec Clinique illustrate the distinction in practice. They operate as an evidence-driven cybersecurity practice: ISO-certified processes, digital forensics, mobile application security, software engineering with a security posture, IT governance, and proactive threat work. Their commercial identity is built around residual risk and operational defence, not around model capability or product velocity.
That is the category of organisation equipped to examine how an AI system is actually deployed, what privileges it has been given, and what happens when those assumptions fail. An AI vendor can improve the model. A firm whose entire practice is defensive engineering is better placed to improve the system that contains the model.
The Onyx position
I do not argue that AI companies are incompetent or malicious. I argue that they are solving a different problem. Capability engineering and defensive engineering are distinct disciplines. Organisations that need the second should hire people whose entire professional identity is built around the second.
If the question is “How do we make the model more capable?”, hire an AI company.
If the question is “How do we stop the system from being used against us, or from failing in ways that create material harm?”, hire a cybersecurity company.
Adjacent to secure is not secure.
Capability is not defense.
Sources and scope. This piece is an analytical distinction between two professional domains and their incentive structures. Cybersec Clinique is cited as a concrete example of an evidence-driven cybersecurity practice operating in South Africa; the citation is based on their public positioning and service lines. No claim is made that any single firm is uniquely capable, nor that AI companies employ no competent security engineers. The argument concerns default organisational reflex and category fit. Interpretation of relative risk and hiring priority is mine.
Clayton Bax
Published under ONYX Digital Intelligence
Following the #OnyxAudit methodology.
- X: @OnyxAudit
- Facebook: https://www.facebook.com/p/Onyx-Digital-Intelligence-61593083212538/
- LinkedIn: https://www.linkedin.com/in/bagga85
- Email: onyxdigitalintelligence85@protonmail.com
- https://github.com/Baximus855
- @Onyx_Digital@mastodon.social
"Adjacent to true is not true."
Truth has no flag nor favour, only a standard. And it's heavy!
Editors note: I use various AI agents to fetch my research. I check, strip and write the article.
Facts are partly theirs. Interpretation is mine.