AI security audit: prompt injection, data leakage and red teaming of agents and chatbots
We attack your chatbot or agent the way a real attacker would and show what worked. After the report we help you fix the findings and test again.
What hurts
Your bot answers customers on your website or in messengers. You do not know whether someone can make it reveal its system prompt, partner pricing or data about other customers.
The agent has access to your CRM, orders or payments. If someone talks it into running an attacker’s command, the damage is no longer reputational but financial.
A standard penetration test covers servers and APIs, not a conversation with a model. An attack through text, a document or an email slips past classic tools.
What we build
- A threat model for your AI system: data, tools and users
- Direct and indirect prompt injection tests, including attacks via documents and websites
- Attempts to extract the system prompt, personal data and commercial data
- Agent tool checks: whether tools can be called with parameters set by an attacker
- Jailbreak, offensive and forbidden answer tests
- Review against the OWASP Top 10 for LLM applications
- A retest after fixes
What you get
- A report with each finding, an attack example and a severity rating
- Fix recommendations for your stack
- An adversarial test suite you can run on every release
- A retest report after fixes
- A walkthrough of results with your team
How we measure the result
- Number of findings by severity
- Share of critical findings closed before the retest
- Number of adversarial tests in the regular run
- Time from a new release to a security check
Systems and stack
Industries
Questions
How much does an AI security audit cost?
From $7,000 over 3–6 weeks. The price depends on the number of channels and agents, what tools the agent has and whether infrastructure testing is needed. An agent with payment access gets a deeper check than a bot answering from a knowledge base.
Will you break our production?
We agree tests in advance and run them in a test environment or during agreed hours. Tools that change data are replaced with mocks. The testing scope is fixed in writing before we start.
Does an audit guarantee the bot cannot be broken?
No, and no honest audit can. Language models can be tricked in new ways that keep appearing. An audit finds known weaknesses, and a regular test suite catches regressions. That is why we recommend pairing it with guardrails.
What do you need from us?
Access to a test environment, a description of the agent tools, the system prompt and the data schema. We sign an NDA before access. Findings go only to agreed people in your company.
Which systems do you test?
Website chatbots, messenger bots on WhatsApp, Telegram and Instagram, voice agents, RAG assistants and internal copilots, on OpenAI, Anthropic, Gemini or self-hosted models. We test the agents on the Happ platform before every release ourselves.
What happens after the audit?
The adversarial test suite stays with you and runs in CI. We can run repeat checks under an SLA when models change or new attack types appear.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.