Hijacking through untrusted content
An agent that reads an e-mail, web page or document can find hidden instructions there and follow them for an attacker, without a person ever validating them.
AI-agent security
We govern AI agents that read, remember and act across your business tools.
Security reviews, controlled red teaming and custom harnesses for organisations whose AI agents access data and make business actions.
What an AI agent can do against you
An agent does more than answer: it reads, remembers, decides and acts inside your tools. Risk is measured by the decisions it can make and their business consequences.
Contaminated memory, skills and automations
A malicious instruction or piece of data can persist in an agent's memory, a skill or a recurring action. The risk then propagates into every new task.
Actions outside business policy
An agent able to send, pay, modify or delete can exceed the rights, thresholds or approvals required by your business process — at a pace no human review can follow.
What we do
Three formats, from a risk decision to controlled agent action in production.
2 days
AI-agent diagnostic
A clear picture: which agents, which access, which risks come first, and where to start.
See details → 5 days · FlagshipAI-agent audit and red teaming
We attack your agents the way an adversary would, within an authorised scope, and show you what gives way.
See details → 10 days or moreHardening, custom harnesses and support
We turn your business rules into verifiable controls, then make them hold in production.
See details →Agent IA Guard
Agent IA Guard: a product direction, controls available now
The product is in development for 2027. Today, we design custom harnesses that apply your business rules before an agent takes action.
Why us
Expertise with three entry points
Mathematics, cybersecurity and artificial intelligence. We understand how a model works, how an attacker reasons and how an organisation decides.
Published research
13 analyses published in French and English on prompt injection, poisoning, agents and hardening, plus a lab of interactive tools.
A referenced method
Tests and recommendations grounded in the field's public frameworks, within a written contractual frame: scope, authorisation, confidentiality, debrief.
Latest publications
All publications →Sep 14, 2026 · 16 min
When the Web Becomes a Weapon: Understanding Attacks Against AI Agents
All it takes to hijack an AI agent is for it to visit a web page carrying invisible instructions. This article walks through the full attack …
Sep 13, 2026 · 18 minThe Hidden Weapon in the Page
An indexed web page can carry a natural-language payload that targets, in a few lines, all four defensive layers of an AI agent: evasion, …
Jul 29, 2026 · 10 minDeceive, Detect, Exhaust: Defending a System Against Offensive AI Agents
A human pentester gets tired, breaks for lunch, eventually gives up. An LLM-driven agent does none of that: it loops around the clock, draws …
Request a security review
Describe the agent, its access and the business decision it can influence. We reply within two business days.
Request a review