Arviser

Arviser

AI Security Review for LLM Apps and Agents

AI security — the AI cybersecurity of systems that read your documents and take actions. Find what your assistant can expose or do beyond its permissions.

Scoped after intake · One application · RAG track available · Warsaw

We test how your AI application handles untrusted instructions, restricted data, and actions that require approval. You get a prioritized list of findings and a plan your engineers can use to fix them.

Checks for prompt injection, data exposure, and excess agent permissions follow risks described in the OWASP Top 10 for LLM Applications (2026) and the OWASP Top 10 for Agentic Applications. See our mapping: OWASP LLM Top 10 2026 — what we check. Quality audits ask how well the system does its job; a security review asks whether someone can make it leak, bypass access rules, or act without approval — compare Security vs Quality Audit.

When to run a review

Run a review before giving an AI assistant access to internal documents, customer records, or tools that can take actions. If the assistant uses retrieval-augmented generation (RAG), treat indexed content as an attack surface: logs, tickets, wiki pages, and vendor docs can carry instructions as easily as facts. If your system is already live, start with the workflows where a mistake would have the greatest impact.

What we test

Prompt injection

Can instructions in a user message, retrieved document, tool response, or log line change how the system behaves?

Data access

Can a user retrieve information through the AI that their role would normally prevent them from seeing — including cross-tenant or cross-team chunks in a shared index?

Agent permissions

Can the agent take an action outside its role or skip a required approval?

RAG track

AI Security Review — RAG track treats the pipeline as one system: sources → ingestion → index → retrieval → prompt assembly → model → tools → rendered answer.

  • Indirect injection — can content in a document, ticket, or error log steer the assistant? See definition.
  • Retrieval ACL — are permissions enforced before chunks enter the model context? See cross-tenant retrieval.
  • Secrets in the index — do configs or logs leak credentials into answers?
  • Output channels — can answers exfiltrate data via links, images, or downstream systems? Related: lethal trifecta.
  • Write path — can a low-privilege editor poison what later users retrieve?

Scenarios are agreed after intake for one application and one critical workflow. We map findings to OWASP LLM categories (and Agentic categories when tools or memory are in scope). This is a point-in-time review, not an OWASP certification.

Illustrative coverage (synthetic)

Example heatmap for a RAG assistant engagement. Real coverage is agreed after intake and listed in the report.

IDScenario familyTypical priority
T-01 / T-02Indirect injection (logs, docs, tickets)P1
T-03Knowledge write-path / poisoningP1
T-04 / T-05Cross-tenant or secret exposureP1
T-06Output / markdown exfil channelP1
T-07Tool misuse / approvals (if tools exist)P1
T-08–T-10Hidden context, vector store, logs/providerP2

Plain-language risks for buyers and product owners: AI security risks (R-01–R-27). Report labels: OWASP LLM Top 10 2026 mapping.

What you receive

  • A short decision brief and a map of data access, identities, tools, and trust boundaries
  • Confirmed findings with severity, reproduction steps, and evidence from the agreed test environment
  • A coverage table of scenarios checked — including what was tested and not found
  • A prioritized fix list your engineers can act on
  • A walkthrough of results; optional bounded retest of agreed fixes

Starting format: review of one application against agreed scenarios. Price and timeline are fixed after we clarify user roles, integrations, and access to a test environment.

Sample finding · Synthetic scenario

A restricted document enters the model context through retrieval. An employee asks a routine question. The retriever includes an excerpt from a file reserved for another team. Even if the model refuses to answer, the retrieval boundary has already failed: restricted text reached the prompt. The recommended fix is to enforce the user’s document permissions before chunks are assembled into the prompt, then re-test with authorized and unauthorized roles (including after access is revoked).

This example is synthetic. For research-style probes on agent permission boundaries, see WardBench (open research on synthetic data).

Review vs security lite

AI Security Review (this page): one application, agreed roles and workflows — including RAG pipelines when in scope — with findings and a fix plan. Price scoped after intake.

LLM security lite: narrow injection and tool-abuse slice for CI (PoC / Eval CI / hardening ladder). Use when you already know the agent path and need reproducible gates — not a full app review.

Need a narrower CI slice? Open LLM security lite

Start with one application

Tell us what your AI system does, what data it can access, what it retrieves, and what actions it can take. A short, non-confidential description is enough. We will agree on the workflows, user roles, and test environment before testing begins.

FAQ

Is this the same as an LLM Quality Audit?

No. A quality audit scores task success, groundedness, and related failure modes. An AI Security Review focuses on injection, restricted data exposure, and agent actions beyond permissions. Full comparison: Security vs Quality Audit.

Do you review RAG / document assistants?

Yes. When retrieval is in scope, we test indirect injection through indexed content, retrieval access control, secrets in the index, output handling, and tool misuse if the assistant can take actions. That is the RAG track.

Is this a penetration test?

No. It is an AI security review of one LLM application. It does not replace classic infrastructure or web/API penetration testing. See AI Security vs pentest.

How is this different from LLM security lite?

Lite is a narrower CI injection and tool-abuse slice with fixed sprint pricing. This Review is a scoped assessment of one application with findings and a fix plan.

Do you publish a fixed from-price?

Not yet. Scope and fee are set after we agree roles, integrations, and test-environment access for one application.

What access do you need?

A non-production or staged environment where we can run agreed scenarios under NDA, ideally with synthetic or canary data. Prefer your stack. No patient data via website forms.

How does WardBench relate?

WardBench is open research on synthetic agent permission scenarios. Client reviews use your workflows and roles; WardBench is cited as methodology context, not as a client deliverable.

Request written scope or ask a question

Free written scoping and technical Q&A reply within 48 hours on workdays (Warsaw), or select Direct Founder Consultation (€250/h, 100% credited in 30 days) below. Controller: Arviser Sp. z o.o. Privacy. No patient data.

1-click prompt templates:
Optional details (role, industry, timezone)

Discuss your project