Once an organization knows which AI services its employees use and has written down what is and is not allowed, the harder part begins: the decision has to be built into the network, the browser, and identity management. How to uncover unapproved use in the first place, and which rules belong in the policy, is covered in our article Shadow AI: when employees use AI without approval. This one walks through the technical tools that can enforce the policy, and how far their reach extends.
These controls work well when employees can get their work done comfortably with the approved tools, while the risky services are either unreachable or usable only in a controlled way.
Why a policy falls short without technical controls
Employees typically read the AI usage policy once, when they join or at a training session. The actual decision gets made every day, when someone facing a deadline pastes part of a contract into a chatbot because it is faster that way. At that moment they rarely remember which service was on the approved list or which account they should sign in with.
This is where technical controls help: the employee gets feedback at the moment of the decision, while there is still time to reconsider.
The other reason is visibility. A policy on its own does not tell you whether people actually follow it. Filtering and logging tools do show which services are in use, and that also gives you a basis for reviewing the policy. Under Article 4 of Regulation (EU) 2024/1689 (the AI Act), organizations that deploy AI systems have been required since February 2, 2025 to take measures supporting the AI literacy of their staff. Training and the policy are part of that, and the technical controls show where it is not working in practice.
Allowed, restricted, and blocked services
Technical enforcement starts with classification. Every known AI service is placed in a category, and each category comes with a technical action. Three levels are enough for most organizations:
| Category | Typical service | Technical action |
|---|---|---|
| Allowed | A tool used under a corporate subscription, with a contract and central administration | Accessible, but only with a corporate account; use is logged |
| Restricted | Useful, but there is no contract for it, or it is suitable only for public data | Warning page, uploads and pasting blocked, or read-only use |
| Blocked | Unknown operator, unclear data handling, free image conversion or translation sites | Blocked on the network and in the browser |
For classification, what counts is what happens to the data you send in. Who operates the service, in which country it stores the data, whether it uses submitted content to further train the model, how long it keeps it, and whether a data processing agreement is available. The same service can fall into more than one category: its enterprise plan is allowed, while the free personal version is blocked.
Many existing applications, from note-taking apps to design software, now have AI features built in. These need to be classified the same way as standalone chatbots, because here too the data usually ends up at an external model. The list needs regular updates, since new services appear every week. For network filtering, it therefore makes sense to build on the AI category maintained by the filtering vendor and to add only exceptions to your own list.
Filtering on the network and in the browser
What the network filter sees
Most AI services are reached over HTTPS, so without decrypting the traffic a network device sees very little of it. At the start of a TLS connection the client usually sends the name of the target server unencrypted (SNI), and the DNS query also reveals the domain. That is enough for a firewall or web filter to block or allow at the domain level. It is not enough to tell a corporate account from a personal one, or to block only file uploads within a service.
Finer-grained rules require the proxy or the NGFW to decrypt TLS. The device then terminates the connection with its own certificate and can see the URL path, the headers, and the uploaded content. For this, the decrypting device's certificate has to be installed on corporate devices, and someone has to decide which categories stay exempt, for example healthcare or banking sites. Decryption has technical limits as well: some applications use built-in certificate pinning and stop working over a decrypted connection.
Ways around the filter
Domain-level filtering can be bypassed in several common ways:
- QUIC (HTTP/3): runs on UDP port 443, and many proxies do not inspect it. If the firewall blocks QUIC for the traffic that should be inspected, the browser falls back to a TCP-based TLS connection, which can be inspected.
- DNS-over-HTTPS: instead of the internal DNS server, the browser sends an encrypted query to an external resolver, so DNS-based filtering never sees the lookup. A browser policy can turn it off or point it to the internal resolver.
- Encrypted Client Hello (ECH): this TLS extension encrypts the server name as well, which leaves SNI-based filtering blind. On managed devices a browser policy can turn it off, as with DNS-over-HTTPS; elsewhere you are left with the decrypting proxy or endpoint filtering.
- Personal VPN, mobile data, home network: if the laptop does not reach the internet through the corporate network, the on-site firewall sees nothing. The answer here is a filtering agent installed on the endpoint, or always-on web filtering delivered from the cloud.
The browser as a control point
Most AI services are used in a browser, so a centrally managed browser covers a lot of what the network cannot see. A policy can specify which extensions may be installed, and that matters because AI extensions can often read the content of every open page. In the corporate browser profile, file uploads or pasting from the clipboard can be blocked on sites in the restricted category, and data loss prevention built into the browser can detect when confidential data in a well-defined format, such as a personal ID number or a payment card number, is about to go into a text field. Which data should stay out of AI tools, and which methods content detection uses, is covered in detail in our article What can go into a prompt: sensitive data and AI tools. Browser-level control only works if employees use the managed profile, so unmanaged browsers and profiles on corporate devices need rules as well.
Corporate or personal account
Enterprise plans typically include single sign-on (SSO), central user management, configurable retention, and an audit log, and the contract often rules out using submitted content to train the model. Which of these actually applies to a given service has to be checked in the contract and the privacy notice. With a personal account, the employee accepts the terms in their own name, and if they leave the company, their conversations leave with them.
A network filter working from the domain cannot tell the two accounts apart, because both load from the same address. Several techniques address this:
- Tenant restrictions: several cloud providers support an HTTP header that the decrypting proxy inserts into the sign-in request and that states which organizational tenants users may sign in to. The provider itself then rejects sign-ins to any other tenant or with a personal account.
- Enforced SSO: accounts in the corporate email domain can sign in only through the organization's identity provider, so conditional access rules, such as requiring MFA and a managed device, apply to the AI service too.
- Rules by browser profile: in the corporate profile, sign-in is possible only with the corporate account; in the personal profile, the service behaves according to the restricted category.
Application consent is worth a look too. Once signed in with a corporate account, many AI tools ask for access to email, the calendar, or file storage. If anyone can grant that, an external service gets read access to the mailbox. The identity platform can be configured so that only an administrator can approve this kind of consent.
Filtering malicious instructions
The controls so far govern who may use which service and what data they may send. For the organization's own AI applications and for assistants wired into workflows, the other direction matters as well: text reaching the model can contain an instruction that hijacks it. This is prompt injection, which the 2025 edition of the OWASP Top 10 for LLM Applications ranks first, as LLM01.
If internal applications call the model providers' APIs through a central AI gateway instead of directly, several checks can run there:
- inspecting incoming prompts and retrieved documents for known injection patterns, using a classifier model and rules;
- detecting and masking confidential data in the text sent to the model;
- checking the response before it goes back to the user or to another system, for example by filtering the domains of external links and images;
- an allowlist of the models and providers that may be used, and central management of API keys so that they do not live in the source code of individual applications.
Filtering reduces the number of successful attacks, but a new phrasing can still get through. That is why, besides the gateway, most of the protection comes from narrowing the tools and permissions attached to the model. If a hijacked model can only read, actions with an external effect need human approval, and links and images in the response are filtered as well, a successful injection does far less damage.
Logging and reporting on AI use
Logging serves two purposes: seeing usage patterns, and tracing events back during an incident. The two need data at different depths. For patterns, metadata is enough: which user, which service, when, how much traffic, and what the filter did (allowed, warned, blocked). You can get this from the logs of the web filter, the proxy, and the identity provider. The audit logs of enterprise AI services add who used which feature and which files they uploaded.
Logging the full text of prompts and responses is a different matter. That text contains employees' personal data, trade secrets, and sometimes very private questions. The GDPR requires a legal basis, purpose limitation, and a defined retention period for it, and under Section 11/A of the Hungarian Labor Code, employees must be told in advance by what means and for what purpose the employer monitors them. A common approach is to record the full text only for the organization's own AI applications and at the gateway, with narrow access and short retention, and to stick to metadata and filter alerts for web use.
It is worth sending these logs to the central SIEM as well, so that AI use can be correlated with other events. Good candidates for alerts include an account uploading many files to an AI service in a short time, someone repeatedly trying to reach a blocked service, or a new AI domain that has not yet been classified showing up in the traffic. The log sources usually needed for incidents are summarized in our article Logging: what to collect for an incident.
Reporting is for decision makers. A monthly or quarterly summary is enough: how many people use which services, where attempts to reach blocked or restricted services have gone up, and which new tools have appeared. When many people try to reach a blocked service, that often means something is missing from the approved tools, and the classification or the procurement needs a second look.
Related service: AI security controls deployment →