← Services

AI cybersecurity

Reducing the risks of your AI systems, without promising the impossible

Gensai makes sure each of its solutions is secured and compliant, from the design stage. When the stakes call for it, a deeper analysis can be carried out, up to penetration tests (pentests). Gensai can also carry out the cybersecurity analysis of an AI application developed in-house or with another provider.

No measure protects an AI system 100%. Language models remain open to manipulation: in February 2026, the UK AI Security Institute presented an automated method that defeated Anthropic's and OpenAI's safeguards, which have been strengthened since. Cybersecurity does not remove the risk, it reduces it: fewer exploitable flaws, limited damage if one is found, and faster detection.

The risks specific to AI come on top of those of any application: prompt injection, direct or hidden in a document or a web page, jailbreaks, leaks of sensitive data or of the system instructions, poisoning of a knowledge base, booby-trapped MCP tools, an agent with too many rights, runaway costs. OWASP lists them in its Top 10 for LLM applications, whose 2026 edition puts prompt injection first and excessive agency third.

The answer lies first in the architecture: rights limited to what is strictly needed, human validation of sensitive actions, outside content treated as untrusted, chosen sources for knowledge bases, logging and usage caps. Then come the tests: the ANSSI, France's cybersecurity agency, recommends security audits before any generative AI system goes into production, from classic penetration tests to attack scenarios run by hand by auditors.

Discuss your AI's security

The types of tests available

Direct prompt injection

Varied phrasings, manual and automated, to make the model ignore its instructions or reveal its system prompt.

Indirect injection

Instructions hidden in a document, an e-mail or a web page read by the AI, to check that it does not carry them out.

Jailbreak

Role play, fictional scenarios, split requests: the known techniques for getting round the model's guardrails.

Data leaks

Checking that a user can obtain neither another user's data nor confidential information from the knowledge base.

RAG poisoning

Measuring the effect of a booby-trapped document in the knowledge base, and checking who can add content to it.

Agents and connected tools

Testing an agent's rights, the MCP servers it uses and what it can do if it is manipulated.

Application and infrastructure

A classic pentest of what surrounds the AI: web interface, API, authentication, servers.

Voice agents

Simulated malicious calls: identity theft, fraudulent requests, attempts to obtain information.

Our references

Frequently asked questions

Does cybersecurity protect an AI system 100%?

No, and no serious provider can promise it. Language models remain vulnerable to attacks that keep evolving. Cybersecurity reduces the risk: it removes the most exploitable flaws, limits the possible damage and makes it possible to detect an incident faster.

What are the main AI-related risks?

Those listed by OWASP in its Top 10 for LLM applications: prompt injection, sensitive information disclosure, supply chain, data poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation and unbounded consumption. Add to these the unsupervised use of AI by employees, known as shadow AI.

How can we protect ourselves?

By combining several protections rather than relying on just one: giving the AI as few rights as possible, having a human validate sensitive actions, treating all outside content as untrusted, choosing and checking the sources of a knowledge base, logging exchanges and capping usage. Then testing, before going into production and after every significant change.

What is an AI pentest?

A penetration test means deliberately attacking a system, with its owner's agreement, to find its flaws before an attacker does. For an AI application, it covers both the attacks specific to models (prompt injection, jailbreak, data leaks, abuse of an agent's tools) and the classic components: web application, API, servers. It ends with a report and recommendations.

Can you test an AI that Gensai did not build?

Yes. Gensai can test AI solutions developed in-house or by another provider, within a scope and a schedule defined with you, and with the written authorisation of the system's owner.

How much does a pentest cost?

There is no official price list. According to a price survey published by Yeeply in September 2026, a basic web application pentest costs between €3,000 and €8,000 in France, over 3 to 5 days, a full web pentest between €8,000 and €25,000, and a red team exercise between €15,000 and €70,000. For an AI application, the price mainly depends on the number of functions to test: chatbot, knowledge base, agent and connected tools. Gensai draws up a quote based on the scope.

What are Gensai's accreditations?

Sébastien Lecocq, co-founder of Gensai, has obtained OpenAI's Daybreak Blue accreditation, which lets vetted cybersecurity professionals use its models with fewer restrictions on security tasks. An equivalent accreditation with Anthropic is under way, expected in early 2027.

Do your tests replace an audit qualified by the ANSSI?

No. Some organisations, such as operators of vital importance or public administrations, must use an audit provider qualified by the ANSSI (PASSI). Gensai does not hold this qualification: its tests then complement the qualified audit, without replacing it.

Is your AI protected enough?

Describe your system, what it does and what it can access: Gensai tells you which tests to run and in what order.

Subject: AI cybersecurity

Information used only to answer the request.