An AI assistant that only writes text and an AI agent that can call APIs, pull sensitive records and start workflows present different security problems. The first can say something wrong. The second can do something wrong.
On September 28, French technology and defence company Thales announced an expanded collaboration with Google Cloud to address that gap.
The companies are integrating Thales AI Security Fabric with Gemini Enterprise, with the stated aim of applying security, governance and visibility across interactions among users, agents, models and tools in real time.
Thales listed the risks the work targets: prompt injection, sensitive data leakage, unsafe outputs, unauthorized actions and increasingly complex agent to agent interactions.
Its own example involves an agent settling insurance claims that exceeds its mandate by using personal information from outside approved sources to influence payouts.
Eva Rudin, Thales’s senior vice president for cybersecurity products, said agents require “a fundamentally different approach to security.” Vineet Bhan, director of security and identity partnerships at Google Cloud, said customers need security and governance built into how agents operate as they gain access to critical data and systems.
The technical claim concerns where enforcement sits. Thales says its controls are applied around agents to keep them within authorized boundaries, limit data access and block inappropriate actions as they happen.
A system prompt or output filter shapes what a model says. An authorization layer decides what an agent is permitted to do, based on the identity acting, the resource requested, the action attempted and the context.
An agent cleared to summarize customer records, for instance, should not inherit the ability to export a whole customer database or change access policies. This example is illustrative and does not come from the Thales announcement.
Broad agent access already has a documented case. In July, Hugging Face reported an intrusion run by an autonomous agent that harvested credentials from a worker’s environment.
Its timeline said the agent went from one worker pod to cluster admin across multiple internal clusters in under thirteen hours.
The release names no customers and gives no availability date. It also includes no performance data.
The Thales announcement does not say how permissions are scoped to individual tasks, whether irreversible actions require human approval, or how tool calls are logged, attributed and revoked.
Those details decide how the controls work in practice, and the release does not publish them.
Featured image: via S3NS
