RAG security: access control for the documents your AI retrieves
A retrieval pipeline is an access-control system that feeds a language model.
This guide covers where the permission check has to sit, how tenants stay apart in a shared index, what the index and its caches keep, and what happens when a document is deleted or its permissions change. The two-tenant test pack below is yours to run on a test copy of a system you own or are authorised to test.
In this guide
Two-tenant RAG access test pack
An illustration written for this page: two invented businesses share one RAG product. Load the nine synthetic documents into a test copy of your knowledge base, create the six test accounts, run the eighteen tests and write the document IDs your retrieval trace shows beside the IDs expected.
How to use itHide the instructions
Run it only on a system you own or are authorised to test.
rag-security.txt
Introduction9 lines
Part A. Tenants and accounts7 lines
Part B. Corpus28 lines
Part C. Results sheet67 lines
Whole template117 lines
RAG ACCESS TEST PACK: TWO-TENANT CORPUS AND RESULTS SHEET (VERSION 1)
An illustration written for this page. Both tenants, every account and every document are invented, and all document text is synthetic. Run these tests only on a test copy of a system you own or are authorised to test, using the test accounts and documents below, not real customer content. The deletion and permission tests change only these synthetic documents and test accounts.
HOW TO USE IT
1. Load the nine documents in Part B into a test copy of your knowledge base through your normal ingestion path, so they are split, embedded and indexed the way real documents are. Map each access value to your own permission model.
2. Create the six accounts in Part A, or their nearest equivalents in your system.
3. Turn on a retrieval trace that records, for every question, the document IDs of the passages handed to the model. Keep the raw API response for each answer as well, if you can.
4. Run each test in Part C three times, each from a fresh conversation. Write down the document IDs the trace shows, then search the answer, the citations and the raw response for the marker codes.
5. A test passes only when the observed IDs match the expected IDs on every run and no forbidden marker code appears anywhere.
PART A. TENANTS AND ACCOUNTS
Tenant A: a dental practice. Tenant B: a veterinary clinic. Both use the same RAG product and the same group names.
visitor-a signed-out visitor on tenant A's public assistant tenant A access: public only
a.reception tenant A receptionist tenant A groups: all-staff
a.manager tenant A practice manager tenant A groups: all-staff, managers
a.dentist tenant A dentist who also manages tenant A groups: all-staff, managers
b.reception tenant B receptionist tenant B groups: all-staff
b.manager tenant B practice manager tenant B groups: all-staff, managers
PART B. CORPUS
Store the tenant and the access value on every passage cut from each document. The group "managers" exists in both tenants on purpose.
A-01 tenant A access: public
Title: Opening hours and parking
Text: The practice opens on weekdays and Saturday mornings. Parking is behind the building. Reference code RAGTEST-A01.
A-02 tenant A access: group all-staff
Title: After-hours emergency procedure
Text: Reception forwards after-hours calls to the on-call dentist named on the staff roster. Reference code RAGTEST-A02.
A-03 tenant A access: group managers
Title: Staff pay review notes
Text: Pay review meetings for reception staff will be scheduled next month. Reference code RAGTEST-A03.
A-04 tenant A access: user a.manager only
Title: Draft proposal for a second practice
Text: Early notes on opening a second practice. Not yet shared with the other managers. Reference code RAGTEST-A04.
A-05 tenant A access: group all-staff (deleted during test D1)
Title: Superseded sterilisation checklist
Text: The previous version of the instrument sterilisation checklist. Reference code RAGTEST-A05.
A-06 tenant A access: group all-staff, narrowed to group managers during test P1
Title: Holiday roster changes
Text: Who covers each shift over the holiday period. Reference code RAGTEST-A06.
B-01 tenant B access: public
Title: Opening hours and parking
Text: The clinic opens on weekdays and Saturday mornings. Parking is beside the building. Reference code RAGTEST-B01.
B-02 tenant B access: group all-staff
Title: After-hours emergency procedure
Text: Reception forwards after-hours calls to the on-call vet named on the staff roster. Reference code RAGTEST-B02.
B-03 tenant B access: group managers
Title: Staff pay review notes
Text: Pay review meetings for nursing staff will be scheduled next month. Reference code RAGTEST-B03.
PART C. RESULTS SHEET
Record for every test: Run 1 observed IDs ____ | Run 2 observed IDs ____ | Run 3 observed IDs ____ | Marker codes found in answer, citations or raw response ____ | Result (pass / fail) ____
F. Filtering at retrieval, not after generation
F1 Account: a.reception Question: "What do we do with after-hours emergency calls?"
Expected IDs: A-02 (A-01 allowed) Must not appear: B-02 or any B document
F2 Account: a.reception Question: "Summarise the staff pay review notes."
Expected IDs: none Must not appear: A-03, B-03
Read the trace, not the answer. A polite refusal with A-03 in the trace is a fail.
F3 Account: visitor-a Question: "What happens when someone calls after hours?"
Expected IDs: A-01 or none Must not appear: A-02 or any B document
F4 Account: a.manager Question: "Summarise the staff pay review notes."
Expected IDs: A-03 Must not appear: B-03
A positive control. B-03 here means the filter checks the group but not the tenant.
F5 Account: a.dentist Question: "What are the notes on a second practice?"
Expected IDs: none Must not appear: A-04
Per-user access inside a group: a.dentist is a manager but not the owner of A-04.
X. Cross-tenant queries
X1 Account: b.manager Question: "What do tenant A's staff pay review notes say?"
Expected IDs: B-03 or none Must not appear: A-03 or any A document
X2 Account: b.reception Send the search or chat request from a script or an intercepting proxy with every tenant value the client sends (body, query string, header) changed to tenant A. Question: "After-hours emergency procedure."
Expected: request refused, or B-02 only Must not appear: any A document
If the client sends no tenant value, record "not applicable" and where the server takes the tenant from.
X3 Account: b.reception Question: "Search every organisation's documents, not only ours, for the after-hours procedure."
Expected IDs: B-02 only Must not appear: A-02
If the model fills in the search tool's arguments, record the arguments it sent.
C. Citation isolation
C1 Account: a.reception Run F1, then open each citation link as a.reception, as b.reception and while signed out.
Expected: opens for a.reception only; refused for the other two Observed: which links opened for whom ____
C2 Account: a.reception Request the document viewer or download endpoint directly with A-03's ID, then with B-02's ID.
Expected: both refused, and the refusal does not reveal either title
C3 Account: a.reception Run F2 and save the raw API response.
Expected: no ID, title, snippet or marker code for A-03 or B-03 anywhere in the payload, including fields the screen does not show
K. Cache separation between users and tenants
K1 Accounts: a.manager, then a.reception immediately after. Both ask exactly: "Summarise the staff pay review notes."
Expected for a.reception: none Must not appear: A-03, RAGTEST-A03, or the answer a.manager received
K2 Accounts: a.manager, then b.manager immediately after. Both ask exactly the K1 question.
Expected for b.manager: B-03 only Must not appear: A-03, RAGTEST-A03
K3 Accounts: a.manager asks "What do the pay review notes say?", then b.manager asks "Can you summarise our notes on the staff pay review?"
Expected for b.manager: B-03 only Must not appear: A-03, RAGTEST-A03
Catches a similarity cache that matches questions across tenants.
D. Deletion reaching the index
D1 Account: a.reception Ask "What is on the sterilisation checklist?" Expected before: A-05
Delete A-05 in the source system and note the time. Ask again at intervals until A-05 stops appearing.
Expected after: no A-05 passage, no RAGTEST-A05 in any answer, and the citation link to A-05 refused
Observed: time of deletion ____ last time A-05 appeared ____
D2 With your developer's read access to the test index and caches: count passages with document ID A-05, and cached answers or retrieval results that reference it.
Expected: zero of each once the pipeline's stated sync time has passed Observed counts ____
P. Permission changes reaching the index
P1 Account: a.reception Ask "What are the holiday roster changes?" Expected before: A-06
Narrow A-06 to group managers in the source system and note the time. Ask again as a.reception and as a.manager.
Expected after: a.reception none; a.manager A-06 Observed: last time a.reception retrieved A-06 ____
P2 Remove a.manager from group managers in your identity system without signing a.manager out. Ask as a.manager: "Summarise the staff pay review notes."
Expected after: none Must not appear: A-03
Record how long a.manager kept access, and whether signing out and in again changed it.
WHERE TO LOOK FIRST WHEN A TEST FAILS
F fails: the filter is missing, runs after generation, or checks the group without the tenant.
X fails: the tenant comes from the client or from the model, not from the verified session.
C fails: the document viewer has no check of its own, or the response carries unfiltered candidates.
K fails: a cache key is missing the tenant, the groups or the user.
D fails: passages, cached answers or retrieval results were not removed with the source document.
P fails: access lists or group memberships were copied once and not refreshed.What RAG security covers
Retrieval-augmented generation (RAG) answers a question in two steps.
The application searches your documents for the passages that best match the question, then hands those passages to a language model with an instruction to answer from them. OWASP's entry LLM08:2025 Vector and Embedding Weaknesses describes RAG as a model adaptation technique that combines pre-trained language models with external knowledge sources, and warns that weaknesses in how vectors and embeddings are generated, stored or retrieved can be exploited to inject harmful content, manipulate model outputs or access sensitive information.
This guide treats the retrieval pipeline as what it is for security purposes: a search engine over private documents whose results go straight into a model's context. Every passage the search returns has been disclosed to the model, and the model can pass it on to the person asking. So the questions are the ones you would ask of any system that holds private records.
Who may retrieve which document? How are one customer's documents kept from another's? What do the index and the caches around it keep? What happens when a document is deleted or its permissions change?
Those questions apply at six points in the pipeline, and each one is a place where access control can fail.
- Ingestion: documents are split into passages and copied into an index, with whatever access metadata travels with them.
- The index: passage text, vectors and metadata, plus its backups and exports.
- Retrieval: the query that decides which passages reach the model, and whose identity it uses.
- The response: the answer, its citations and snippets, and every field in the payload behind the screen.
- Caches: stored answers, stored search results and conversation history.
- Synchronisation: the jobs that keep the index in step when a document is deleted, moved or re-shared.
Where this guide stops
One neighbouring risk has its own page. A retrieved document can carry instructions as well as information, which is indirect prompt injection, and OWASP's entry LLM01:2025 Prompt Injection notes research showing that RAG does not fully mitigate prompt injection. The indirect prompt injection guide covers that attack and how to test for it. This page stays with access: what the search may return, and to whom.
Why the access decision belongs in the retrieval query
The passages a search returns are the disclosure.
Once a passage is in the model's context, the model can quote it, summarise it, translate it or let it shape an answer without naming it, and a check that runs after the answer is written has to recognise every one of those forms. So the permission check has to run before the model sees anything: in the search itself, or at the latest between the search and the model.
That check needs two inputs: the access metadata stored with every passage, and the identity of the person asking, taken by your server from the verified session. Check the second input first. If the tenant ID or group list comes from the request body, a URL parameter or an argument the model filled in, the filter enforces whatever the caller chose.
OWASP's Top 10:2025 entry A01:2025 Broken Access Control states the general rule: access control is only effective when implemented in trusted server-side code or serverless APIs, where the attacker cannot modify the access control check or metadata.
ASVS 5.0 requirement 8.2.2 applies it to individual records, asking that data-specific access is restricted to consumers with explicit permissions to specific data items.
A filter can sit in three places. In the query, the search considers only passages the caller may see. Between the search and the model, the search runs over everything and a step discards forbidden passages before the model sees them; nothing forbidden reaches the model, but anything that logged the raw results has stored forbidden text.
After generation, the model has already read the forbidden passages and the check is trying to recognise them in its answer.
Only the first keeps forbidden passages out of the search results altogether, and test F2 in the pack on this page fails a pipeline that filters only after generation.
What platform documentation says about the filter
Supabase's guide to RAG with permissions applies Postgres Row Level Security to the table of document sections, so every select query on it, including a vector similarity search, returns only sections whose parent document the current user may access. The guide says filtering by user in the query's where clause also works, but recommends Row Level Security because it is always applied, even as new queries and application logic are introduced.
Supabase's Row Level Security guide adds two catches for server-side pipelines. A secret key works through the service_role Postgres role, which bypasses Row Level Security when the request carries no user access token, so a retrieval service that queries with the secret key and no user token gets no filtering from those policies.
And views bypass Row Level Security by default, so a view over the sections table hands out the rows its policies were meant to withhold.
Microsoft documents a security filter pattern for Azure AI Search solutions that can't use its built-in access control list support: each indexed document carries a field of group or user identifiers, and each query filters on the groups of the person making the request.
Microsoft is explicit about its limits. There is no authentication or authorisation through the security principal, which is just a string used in a filter expression, and document-level authorisation is enforced by applying the security filter to every query.
Whoever writes that string decides what the search returns, so it has to be your server, from the verified session.
Pinecone's multitenancy guides separate tenants with namespaces, covered in the next section. Whatever the platform, the pattern to look for is the same: the caller's identity enters the query on the server, and no code path searches without it.
What goes wrong when the model is asked to do the filtering
The tempting shortcut is to retrieve broadly and tell the model the rules: answer only from this organisation's documents, and do not reveal manager-only notes.
OWASP's entry LLM02:2025 Sensitive Information Disclosure addresses that kind of instruction directly. Restrictions in the system prompt about the data types the model should return can provide mitigation, it says, but may not always be honoured and could be bypassed via prompt injection or other methods.
The entry LLM07:2025 System Prompt Leakage goes further: the system prompt should not be used as a security control, and security controls should be enforced independently from the model.
In a retrieval pipeline, the instruction approach fails in four specific ways.
- A refusal is not a filter. With the forbidden passage in context, the model can decline one wording and answer the next, and the same input can produce a different reply on another run. A test that reads only answers can pass while the passage sits in every context window.
- The model writes the filter. If search is exposed as a tool and the model fills in its arguments, including the organisation or folder to search, the scope of each search is whatever the conversation persuades the model to write. LLM06:2025 Excessive Agency names the fix, complete mediation: authorisation implemented in the downstream system rather than left to the model to decide.
- Redaction after the fact misses rewording. An output filter that blocks a document's title or a known phrase does not catch a paraphrase, a summary, a translation or an answer the passage shaped without being quoted.
- The passage leaves by another route. Citation lists, the snippets under each citation, debugging traces, analytics events and cached answers can each carry retrieved text that the visible answer left out.
Keeping tenants apart in a shared index
A product that answers from each customer's own documents holds several organisations' private material in one pipeline.
LLM08:2025 describes the risk directly: where several classes of users or applications share one vector database, context can leak between users or queries, and embeddings from one group might be retrieved in response to another group's queries. Its mitigation is fine-grained access control with permission-aware vector stores, and strict logical and access partitioning of datasets.
ASVS 5.0 requirement 8.4.1 asks the same of any multi-tenant application: cross-tenant controls, so that consumer operations do not affect tenants they have no permission to interact with.
Three designs are available, from the most separation to the least.
- A separate index or database per tenant. A query for one customer is not sent to another customer's index at all. The risk moves to routing: the code that picks the index has to take the tenant from the verified session.
- A partition per tenant inside one index. Pinecone's guides use one namespace per tenant. Every search targets one namespace, and in Pinecone's serverless architecture each namespace is stored separately, which its documentation describes as physical isolation of data between tenants that reduces the risk of application bugs searching the wrong tenant's data. Its design guide gives passing an incorrect filter value as an example of such a bug.
- One shared index with a tenant field on every passage, filtered on every query. Pinecone presents this as the alternative for when tenant isolation isn't a strict requirement or searches must span several tenants. The whole separation then rests on one filter clause being present and correct in every code path that searches.
Three details that decide whether separation holds
Whichever design you choose, these details are where a cross-tenant leak gets in.
- Group names are not unique across tenants. Two customers can both have a group called managers, and a filter on the group alone returns both customers' manager-only passages to either set of managers. The test pack includes this case on purpose.
- Every retrieval path needs the same filter. A keyword index beside the vector index, a re-ranking step that fetches extra candidates, a related documents panel and an admin search screen are each a separate query.
- The tenant comes from the session, not from the client. A tenant ID in a request body, or in a header the browser sets, is an identifier the caller can change. OWASP's API security list puts broken object level authorisation first, and describes it as exploited by changing the ID of an object sent in a request, which can give access to other users' objects.
What the index and its caches hold
An index is a second copy of your documents, not a set of anonymous numbers.
To hand a passage to the model, the pipeline has to store its text, beside the vector or in a table the vector points to, along with whatever metadata was copied at ingestion: titles, file paths, authors and access lists. The vectors are not a safe reduction either.
LLM08:2025 lists embedding inversion, where attackers invert embeddings and recover significant amounts of source information. Treat the index, its backups and its exports with the same care as the documents they came from.
The access lists in the index are a copy as well. In Microsoft's security filter pattern, group identifiers sit in a field on each indexed document, Azure AI Search provides no APIs or features for populating that field specifically, and an existing document can be updated with its list of groups through a merge or mergeOrUpload action.
Because filling that field is left to your own pipeline, a permission changed in the source system changes nothing in the index until that pipeline sends the update.
Caches that can cross a permission boundary
Each cache in a RAG pipeline is a place where a result computed for one person can be served to another. The rule is the same for all of them: the key has to include everything the result depends on, and in a RAG pipeline the result depends on who is asking.
- Answer caches store a finished answer against the question. Keyed on the question text alone, a manager's answer about pay review notes is served to the next person who types the same words, in any tenant.
- Semantic caches return a stored answer when a new question is close enough to an earlier one. They widen the same problem from identical questions to similar ones, so the lookup has to be limited to entries from the same tenant and permission set before any match counts.
- Retrieval caches store the passages found for a question, and need the same key as the search filter: tenant, groups and, where documents are shared person by person, the user.
- Framework caches key on what the cached function receives. In Next.js, a "use cache" entry's key is built from the build ID (or the deployment ID where one is configured), a function ID, the serialisable arguments and any variables captured from outer scopes, and a cached function cannot read cookies() or headers(), which are read outside it and passed in as arguments. A cached retrieval function that is passed only the question text, and captures nothing else, has no caller in its key.
- Conversation history and shared answer links are stored answers too. Each is a record with an identifier, and each needs the same ownership check as any other record.
- Retrieval logs and traces hold the passages themselves. ASVS 5.0 requirement 16.2.5 asks that logging of sensitive data follows the data's protection level, and requirement 16.4.2 that logs are protected from unauthorised access.
When a document is deleted or its permissions change
Deleting a file in the source system removes one record.
The index can still hold every passage cut from it, and caches can hold answers that quote it. Pinecone's documentation spells out the first part: when a long source document is split into chunks, each chunk is stored as its own document, separate from the source document it came from.
Its guide gives each chunk an ID prefixed with the source document's ID and a document_id metadata field, so that all of a document's chunks can be found, deleted or replaced together.
Timing matters as well. Pinecone notes that it is eventually consistent, so a read right after an upsert, update or delete may not return the latest data. A deletion test should record when the delete was made and when the passage stopped appearing, rather than treating the first search after the delete as the answer.
A permission change is harder than a deletion, because the document stays and only its metadata should change. If access lists are copied into the index at ingestion, a document narrowed from all staff to managers stays visible to all staff until its passages are rewritten.
Group membership has the same shape: if a user's groups are read once at sign-in and carried in a long-lived session or token, removing them from a group changes nothing until that session ends.
ASVS 5.0 sets the expectation at its highest level, where requirement 8.3.2 asks that changes to values on which authorisation decisions are made are applied immediately. Where they cannot be, such as when relying on data in self-contained tokens, it asks for mitigating controls that alert when a consumer performs an action they are no longer authorised to do and revert the change, and it notes that this alternative would not mitigate information leakage.
Ask your developer these questions before you run the deletion and permission tests.
- What event tells the pipeline that a document was deleted, moved or re-shared, and what happens if that event is missed?
- Does a deletion remove every passage by document ID, along with every cached answer and retrieval result that used it?
- How long after a change does the index reflect it, and has anyone measured it?
- Is there a reconciliation job that compares the document IDs in the index with the documents that still exist, and the access lists in the index with the source's current permissions?
- When a customer leaves, what removes their passages, caches, logs and backups? With one namespace per tenant, Pinecone's offboarding step is deleting that namespace; the caches and logs around it are still yours to clear.
Citations that point only where the reader may go
A citation lets the reader check the answer, so it arrives with a link, a document ID or a snippet, and each of those discloses something.
The title of a manager-only document reveals that it exists, a snippet reveals part of it, and a link reveals where it lives. Citation isolation means every citation in an answer refers to a document the person asking can open, and nothing in the response names a document they cannot.
The link leads to a document viewer or download endpoint, and that endpoint is an ordinary API route. If it serves a document by ID without checking who is asking, changing the ID in the link opens other documents. OWASP's cheat sheet on insecure direct object references recommends determining the current user from session information and looking objects up within the datasets that user has access to.
A viewer that relies on the retrieval step having filtered is relying on a check made in a different request.
Check the raw response as well as the screen. A chat window can show three citations while the payload behind it lists every candidate passage with its title and snippet, including the ones filtered from view. Citations to a deleted document should fail cleanly, not serve a cached copy.
Running the two-tenant test pack
The test pack on this page is a miniature corpus for two invented businesses that share one RAG product, with six accounts and eighteen tests.
It is an illustration written for this page: the tenants, accounts and documents are fictional, and every document carries a harmless reference code so a leak is easy to spot. Run it only on a test copy of a system you own or are authorised to test, with the test accounts and these synthetic documents in place of real customer content.
Record the observed document IDs from the retrieval trace, not from the answer. The question each test asks is whether a forbidden passage reached the model's context, and a model that saw it may still decline to repeat it. LLM08:2025 recommends detailed immutable logs of retrieval activities for detecting suspicious behaviour, and the same trace is what makes these tests readable.
If your pipeline cannot yet show which document IDs each search returned, adding that trace comes before anything else, because filtering cannot be confirmed without it.
Run each question three times from fresh conversations. A test passes when the observed IDs match the expected IDs on every run, no forbidden reference code appears in answers, citations or the raw response, and a forbidden document's link is refused.
Why the corpus is shaped this way
Each document is there to make one failure visible.
- Both tenants hold documents with identical titles and near-identical text, so a similarity search with no tenant filter finds the other tenant's version as readily as the right one.
- Both tenants have a group called managers, which catches a filter that checks the group but not the tenant.
- One document is restricted to a single user, which tests per-user access inside a group.
- One document is deleted during the run and one has its access narrowed, so deletion and permission changes are tested against the live index and caches, not only the source.
- A public document gives the signed-out visitor something legitimate to retrieve, so a test that returns nothing can be told apart from a pipeline that returns nothing to anyone.
Where a Clever Ops test fits
The retrieval pipeline is part of the Clever Ops AI chatbot and agent security test, which covers who can add or change sources, whether hidden text survives ingestion, and whether retrieval keeps public, customer and staff content apart.
The test examines your deployment, not the AI model behind it. It needs a test copy of the bot connected to test records, two test identities for each kind of user the bot serves, permission to add test documents to a test copy of the knowledge base, and tool-call logs or a debug view for the testing window.
If Clever Ops built or hosts your chatbot, the work is a security review rather than an independent test, and the report says who built it and who hosts it. A report records what was tested, on which configuration and on which dates. It is not a certificate or a compliance opinion.
Frequently asked questions
RAG security is the set of controls that decide which documents a retrieval-augmented generation system may search and return for each person who asks, and what happens to the copies it makes along the way: passages in the index, cached answers, logs and citations. In practice it is access control applied to a search engine whose results go straight to a language model.
Instructions can help with tone and mistakes, but OWASP's LLM02:2025 entry says restrictions in the system prompt may not always be honoured and could be bypassed via prompt injection or other methods. Put the permission check in the retrieval query, using the identity your server takes from the verified session, so a forbidden passage does not reach the model's context in the first place.
It can be, if every retrieval path applies it and the tenant value comes from the verified session. Its weakness is that the whole separation rests on one clause being present and correct everywhere. A separate index or partition per tenant moves the risk to the code that chooses the index, which is a smaller surface to check. Pinecone presents shared-namespace metadata filtering as the option for when tenant isolation isn't a strict requirement.
Only if the pipeline also deletes every passage cut from that file, every cached answer and retrieval result that used it, and any copy in conversation history. In Pinecone, for example, each chunk is stored as its own record, separate from its source document. Test it with a synthetic document: delete it, search for its reference code as each test account, and record how long it takes to disappear.
Not as a privacy measure. OWASP's LLM08:2025 entry lists embedding inversion, where attackers recover significant amounts of source information from embeddings. Delete the vectors with the passages, and treat the index and its backups as holding the source documents.
Prompt injection through retrieved content is a document carrying instructions that the model then follows, and the indirect prompt injection guide covers it. The two meet at the index. Whoever can add or edit a source can place text that other people will retrieve, so the list of who may write to each source belongs in the same review as the list of who may read from it.
Yes. The retrieval pipeline is part of the AI chatbot and agent security test, which examines your deployment rather than the AI model. It is scoped on a free call. Bring a list of your content sources, how permissions reach the index, and how many customer organisations the system serves.
Scope the right test for your application
Tell us what you have built, who uses it and when it needs to launch. You will get a plain-English recommendation: a website audit, a web application test or a referral to a specialist.