Meetup AWS UG: agents IA, identité et conformité

Inside the event

Hello everyone 👋 🚀 The AWS User Group Paris is delighted to welcome you to a new meetup, this time at Lenstra's offices! An evening devoted to a question few teams have truly settled: how to keep control over AI agents that decide on their own what to do, and how to prove it to an auditor. \*\*\* **📆 Thursday, October 15, 2026** **📍 Lenstra's offices, 34 avenue des Champs-Élysées, 75008 Paris** **🕡 Doors open: 6:30 PM \| Talks start: 7:00 PM** **🍕 Buffet offered by Lenstra at the end of the evening** \*\*\* **On the agenda 🍿** **1️⃣ Talk 1 by Lenstra: “AI agents on AWS: who is allowed to do what, and how do you prove it?”, by Thomas Mattone, Forward Deployed AI Engineer at Lenstra** AI agents don't follow a pre-written workflow: it's the model that decides, in real time, which tools to call and in what order. They read sensitive data, write to your systems of record (CRM, core banking, ERP, ticketing), and do so either with their own identity or on behalf of the human they are interacting with. One question remains that few teams have truly settled: who is allowed to do what, with which identity, and how do you demonstrate it to an auditor? The Lenstra team will begin with its convictions: caps and guardrails enforced by the platform, not guidelines written in a wiki that nobody reads. Simple decision rules for teams building agents: when should an agent be given its own authority? When should it act on behalf of the user? And how to decide case by case without slowing down deliveries? In short, traceability of every action, attributed to an identity, that holds up against today's and tomorrow's regulations. You will then discover a real-world implementation on AWS. With Amazon Bedrock AgentCore Identity, both identity modes are supported: the agent's own, or that of the user on whose behalf it acts. AgentCore Gateway then exposes the tools behind a single access control point, regardless of the mode used. When the agent acts with its own authority, IAM sets the limit it cannot cross, no matter what it decides. When it acts on behalf of a user, scoped tokens carry that same limit. Traceability remains: CloudTrail provides the audit trail, while observability traces and logs, pushed to Dynatrace, make it possible to track what each agent actually did. This talk is aimed at architects, tech leads, and security leads who want to run AI agents at scale without ever losing control. **2️⃣ Talk 2 by Cecurity: “Implementing responsible AI on AWS: from the AI Act to technical guardrails”, by Quentin Cozette, Head of Governance, Risk & Compliance / DPO at Cecurity** Adopting generative AI requires meeting regulatory requirements (AI Act, GDPR, ISO 42001) while keeping security and governance in the cloud under control. Drawing on his role as Head of Governance, Risk & Compliance / DPO at Cecurity.com, Quentin will show how to translate these obligations into technical controls on AWS: designing guardrails with Amazon Bedrock Guardrails, managing confidentiality and traceability, and using IAM, CloudTrail, and CloudWatch to document and audit AI workloads. The session is aimed at security, data, ML, and legal teams who want to deploy compliant, safe, and operational AI applications. \*\*\* 🙏 Many thanks to Lenstra for hosting us in their offices and for the buffet that will close the evening. ⚠️ Limited seats: register quickly, and remember to update your registration if you end up unable to attend, so as to free up seats for other members. We look forward to seeing many of you on October 15! 🚀 The AWS User Group France - Paris team