Agent IA Guard — in development — expected in 2027
An AI agent in production obeys what it reads, not only what it is asked. Instructions placed inside the model are not an access control: they can be bypassed by the content the agent processes.
Control principle
Agent IA Guard aims to place a control layer between the agent and its tools, outside the model: every sensitive action is assessed against explicit rights, business rules, thresholds and approvals before it runs, whatever instructions the agent has received.
Who it is for
Companies and institutions that run, or plan to run, AI agents connected to business tools, and want a say in how the product is shaped.
What is being validated
- The range of actions, data and tools the control layer can cover.
- How to express business rights and policies without adding burden on teams.
- The impact on latency and on the agent's user experience.
- Integration paths with the most common agent platforms.
Pilot partner programme
Pilot partners get access to working versions, steer priorities and benefit from preferential terms at launch. Places are limited.
The information on this page will evolve as development progresses.