AI agent security audit

Our AI employee qualified as an “AI Security Auditor” tests chatbots, agents and RAG for prompt injection, data leaks and excessive tool permissions. You receive a Google Doc with findings, evidence and a fix plan. The specialist is available online in your panel with the Max plan or Max Specialists add-on ($35/month).

$50price
2–3 daysturnaround
OWASP Top 10 for LLMstandard
Order an auditWhat we check
Order

Order an audit

We reply within a business day

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.

Need more than a security audit?

We also offer an AI agent architecture audit for $100: request path, models and providers, prompts, RAG, tools, resilience and cost. C4 diagrams and a ranked change plan, 2–5 days.

About the architecture audit

What we check

  • Prompt injection — direct and through documents, emails, web pages and attachments the agent reads. Attempts to rewrite instructions, extract the system prompt, make the agent act on someone else's behalf.
  • Data leaks — other users' conversations, personal data, secrets and keys in answers, logs and traces; tenant isolation in multi-tenant systems.
  • Excessive agency — what the agent can do with tools and APIs: send, delete, pay, run commands, reach internal addresses (SSRF). We verify confirmations and trust boundaries.
  • RAG and the knowledge base — poisoning through uploaded documents and sites, injections in retrieved chunks, leaks through search over other tenants' data.
  • Insecure output handling — HTML and scripts in answers, phishing links, execution of generated code.
  • Abuse and cost — missing limits, token burning, denial of service through long inputs, bypassing paid restrictions.
  • Supply chain — model providers, plugins, MCP servers, third-party packages and their permissions.

We use STRIDE to model threats across data flows and trust boundaries, then test the system against the OWASP Top 10 for LLM Applications and OWASP agent security guidance. We extend this with our own dialogue attack set: 30+ scenarios we run daily against our own AI employees.

How the audit goes

  • Recon. Access to a test environment or the live system with a test account, reading code and prompts if you are willing to share them. Without code we work black-box.
  • Attacks. Manual and automated dialogue probes, checks of tools, permissions, logs and storage. Only agreed scenarios, no load on production and no actions with real money.
  • Report. A coverage table: every checked case marked OK or FAIL; every failure with reproduction, impact, severity and a concrete fix. Separately: what was not checked and why.
  • Re-check. After your fixes we rerun the failed scenarios and confirm closure.

What you get

  • A report for the executive and for developers: short conclusions, risk matrix, detailed findings.
  • A set of test dialogues you can put into CI and run before every release.
  • A prioritised fix plan with effort estimates.

2–3 days. Fixed price: $50.