AI security starts at the network and does not end there
A firewall is the first of seven security layers around an AI system. On-premise hosting fixes that layer and helps with part of the data layer. The other five need the same work as in any production system.
ENISA's Threat Landscape 2025, published on 1 October 2025, analysed “4875 incidents over a period spanning from 1 July 2024 to 30 June 2025”. It also recorded how quickly the attackers adopted the technology: “By early 2025, AI-supported phishing campaigns reportedly represented more than 80 percent of observed social engineering activity worldwide.”
Juhan Lepassaar, ENISA's executive director, described the structural risk: “Systems and services that we rely on in our daily lives are intertwined, so a disruption on one end can have a ripple effect across the supply chain.” An AI assistant is one of those systems. It has access to documents and, increasingly, to other systems.
The seven layers, and what each one decides
- Network. Decide what the system may talk to, in both directions. On-premise, outbound access can be removed rather than filtered.
- Identity. Know which human is behind each request. Shared accounts on an assistant destroy every audit trail underneath them.
- Authorisation. Know what that human may reach, and enforce it where the documents are retrieved rather than where the answer is written.
- Data. Classify what the system is allowed to process, and treat the index and the logs as copies of the source, with the same protection.
- Application. Validate what goes in and what comes out. Generated text that becomes a query, a command or an email is untrusted input.
- Model. Know which version is running, where it came from and what it is weak at. OWASP's 2025 list has supply chain at LLM03 and misinformation at LLM09.
- Operations. Patch it, monitor it, test it and be able to recover it.
What local deployment adds, and what it does not replace
On-premise hosting fixes layer one and helps with part of layer four. It gives the organisation direct control of the inference environment: which model, which version, which network path, which update. A broader claim does not survive the first security review.
For layers six and seven the location gives only the choice of model, version and update: the vetting and the patching stay to be done. For layers two, three and five it does nothing. It does not replace identity management, endpoint security, application security or governance, and the system still has to be secured the way any production system is secured.
With Bastion
What Bastion changes
Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.
For the unit it ships, Bastion owns layer one and also covers layers six and seven: no outbound connection required, a known model and version, and a signed update path with whole-unit replacement.
Identity, authorisation, data classification and output review belong to the customer's environment.
Questions this article answers
- What are the main AI security risks for a company?
- They sit in the application, not the network: prompt injection, disclosure of sensitive information, a compromised supply chain, poisoned data, unvalidated outputs, excessive permissions and a model nobody is monitoring. ENISA also recorded AI-supported phishing at more than 80% of observed social engineering by early 2025.
- Is putting AI behind a firewall enough?
- No: network control is one of seven layers. Identity, authorisation, data classification, application validation, model provenance and operations decide most of the outcome, and none of them changes because the server moved.
- What does on-premise AI actually improve?
- One layer and part of another: the network, because outbound access can be removed rather than filtered, and the data layer, because the index and logs stay inside. It does not replace identity management, endpoint security or governance.
On the record
- 1
- 2
Read next
- The OWASP Top 10 for LLM applications, read as a buyer's checklistThe ten risks in plain business language, which of them an on-premise deployment touches, which it does not, and the minimum an organisation of thirty people should do about them.
- Prompt injection: the attack that needs no clickLanguage models do not separate instructions from content, so any text the model reads can behave like a command. What EchoLeak showed, why retrieval makes it worse, and why running the model locally does not fix it.