# Building Frontier AI — Lélio Renard Lavaud — session 2026-09-23T16:40:00.000Z → 2026-09-23T17:05:00.000Z _77 transcript lines · 46 slides_ ## Transcript It's not a bunch of, like, things that you've seen. It actually organizes all my projects in a nice way. So I can go let's say, I'm gonna go to Project. I'm gonna go in here. And here I have a bunch of to dos that I wanna do. I go here. There's like a plants and heartbeats and a bunch of things, and there's a factory, which if I click, you're gonna see these are the processes that every task goes to. So if I add a task, we go through the sole bot, which is like coding on the thing. Then we have so fast, which is reviewing the thing, and they're going in a cycle until it's like fully reviewed. And then when it's fully reviewed, we have the tester, which clicks around and checks if the thing is done. So if I go back from the factory to one of the projects, and if I wanna get something Done. I just go in the to do here, and I can say I don't know. Now do the thing again. If I press enter, this won't work right now. But it should jump into ready, and it should go through the entire factory process until it's in review. And then you get, like, a beautiful interface where you review it, and you actually feel like you're coordinating a factory and not just talking to a bloated orchestrator app. I am out of time. I wanted to show you this more. It's called Bench. It's gonna be announced on my Twitter. Hopefully, bully me, please, soon. Because it really changed the way I don't use CLIs. I don't use codec apps. I don't use anything. I can Open it on my phone. I can open it anywhere. And just from any underpowered device, I can orchestrate my entire slop factory of things. So thank you for listening. I'm out of time. I have a flight to catch, and appreciate you listening to my rambles. Thank you. Thank you. Let's give it up one. Thank you. Alright. Let's give it up one more time for the Kitsap, please. Wow. I forgot to Say that Nikhite is one of, actually the most viewed speaker, the AI engineer. Like, if you go on AI engineer on YouTube, you'll see so many of his talks, and they're, like, they rank always among the the best and the most viewed. So, it was also interesting to see how, I relate to many of those problems in my day to day. I forgot to say that I work for Replit, and one of the things that we do is we we we we build, you know, infrastructure to to to, to to, build applications as well. Sorry. I'm gonna get Alright. At some point. Alright. So we are going to still talk about the future. And now we're gonna hear from VP of Engineering of Mistral, and he's also, the third employee of the company. So please, join me in welcoming to the stage, Nelio Hello, everyone. Thank you so much for, coming here tonight, for being so many, today. We are gonna talk a bit about what Mistral has been cooking, what's new, and hopefully give you a few information about what we are, actually been working on. So for the ones who do not know us, we are Mistral. We are The top frontier AI lab in Europe. We started, three years and four months ago. Very small company at the time. It's now 1,300 of us. So we grew a lot. We have a clear mission, putting AI in the hands of customers, changing the way enterprises operate. So we're not looking for AGI. We're looking for usefulness, and this is why we now have a team that is dedicated in making things, underground, turning POCs into production, and this is, basically what the Is doing now. So if we look back at the previous AI Engineer Summit a year ago, I was, very excited to be invited last year by. Some things happened, since then. You are more than 700 to be registered. We have more than 50 speakers, today, so it's an amazing event. And about COEB, that was inviting us, last year, we now have the chance of having them within Mistral. And you could ask Why did we make this acquisition? Well, as we are moving towards agentic workloads, swarm of agents going everywhere, it made so much sense to get a team that focused on building this kind of agentic workloads, completely different from getting servers and things like this, just focusing on the sandbox, focusing on the unit of compute that you need to actually run workloads at scale. This is what Jan said, last year, if it works. From day one, we've stayed close to the hardware to provide high performance with sustainable costs. We are operating tens of thousands of applications of on top of bare metal servers across 10 locations worldwide, and they're joining Mistral to build that infrastructure. So making sure that for our own workloads, anything that we need to build and that requires secure, safe, and fast async execution in the background, we could leverage their technology. We've all So in May, announced the acquisition of Emmy AI. Emmy is a company that is focused on physical AI simulation. And if I read the same quote from Johannes, by integrating our expertise into mutual AI's world class AI ecosystem, we're positioned to revolutionize core r and d. Together, we are providing the foundational intelligence required to design and build the next generation of aircraft, vehicles, and semiconductors. So this is really about bringing all the capabilities, the simulation. All the things that actually help understand and model the word back into the models so that we can actually help our industry customers be better. We announced earlier this year, a $150,000,000 partnership with Orbis, and this is typically where having this kind of expertise in house will help us build state of the art specialized models, for this kind of customers. We also announced, last week our series d. This is the biggest From a technical company, in Europe, so €3,000,000,000. We are very proud that this is also gathering a lot of investors from Europe, from all over the world who trust us with this mission of actually making AI useful, of transforming, companies, and this is showing that, the way we've been working, for the past three years is actually working. But I... I'll do it again. You you watched me. I didn't do it. If it happens next, you... You'll know that it's me. I'm gonna talk briefly about the future of work and the way we see it. This means how we also deal with agentic safety. We're gonna talk about sovereignty a little bit and then how we are, working towards new models. I know you all came to hear About. I'm not gonna announce it today, but, rest assured that we're working hard to get it there. So, this is a tough one, apparently so. Okay. Perfect. So the future of work, let me talk a bit, about our product suite today and how we envision it. We, do have a strong product for developers called the Studio. In which we build a lot of primitives, on top of the models themselves. So you have so many models today, and that's not enough to go into production. And that's why we developed the studio, which brings a ton of product value on top of this, things that you actually need to go into production. So think about workflows like long lived execution, the sandboxes that I mentioned earlier, all the primitives that you need to actually make this happen. And we do have Vibe, that is our end user facing product that leverages all those capabilities, to act Actually, well, solve problems for productivity for, end user customers, and especially enterprise users. Obviously, we also have a strong solutions team that is building amazing use cases for all our customers, and we are also having ops science team, obviously, building models, generic models, but also specialized models for everyone. But what we really wanna get to, since I have no power here, sorry, is there. So this is Steven. Steven is Nice person. He works in finance, at Mistral, and he is in charge of making sure our revenue is properly used. Steven is what some could call a tinkerer, but he doesn't have time to build, new workstations in his basement, unfortunately. So he needs to have a tool that will allow him to ask very simple questions, but also build applications. He needs to have access to tools that will let him, build, for instance, Application to analyze and cross check some data coming from the data lake, from the Excel spreadsheet, from the, Notion, workspace, and make sure that he can wrap that up into an app that he can share to, basically his coworkers. So this is pretty much what we heard before, but he operates in a corporate environment where there's some guidelines where we need to make sure that the data is properly used. And this is the unified version of Vibe that we're, actually working on is how we make sure that this kind of user Of builders who are just regular end users, but starting to know more and more about applications, about how to use those models to create useful stuff are gonna leverage, our product suite to do so. So this is typically an automation that we've been working on, mostly for the engineers, how we start from a Slack alert that something is wrong in the middle of the night to an agent investigation. So making sure that basically the on call engineer that is getting paged While he wakes up, while he goes to his computer, has ready to go information to address, when they try to address and process the alert. And this is working typically with, our Vibe stack because we'll have immediately asynchronous agents running on some boxes. You got it. That will plug to the connector that we have, graph connector, centric connectors, GitHub as well to understand what could have regressed all of that and propose, some investigation. Some clues that will help the engineers that is on call, among a huge system to understand where the issue could be. And then obviously post back the findings on Slack, so it can be a conversation. And so the on call engineers are able to start from somewhere without wasting too much time. So this is, an example. You'll have the initial Sentry alert, and immediately the agent is answering with some of the things, that, it found them. That will help the engineer understand what's going on. What's funny with this is that you used to need to do a lot of plumbing to do this because you're plugging multiple systems. You need the, agent to be running somewhere. You need to be setting up those connections, saying you need access to this connector and to this other connector. And this is why we're focusing on this, unified vibe version is where you will start from this kind of prompt that is simple enough for our dear Steven, to enter it. When interview feedback is missing, send A personalized reminder to the interviewer who who forgot to do it. And before you need to think about connecting the right FCPE of your ATS or something, what we're working on is making sure that this is pretty much everything that you have to do. Because all we've just seen in the previous presentation is entirely right, and you need to think about it. But for our own tinkerers, enterprise tinkerers, who are actually not developers and who actually know nothing about Linux VMs or something. They just wanna be able to Send such a prompt and have all the connectors and all the enterprise tools that they need readily available. And this is where Vibe can now understand this intention, spawn agents in the background, build some code, run it, run the test, deploy them as well, host them, and basically iterate by itself for minutes or hours till you have an end result that you can actually share to your coworkers with a set of, you know, restrictive, access controls based Your permissions, all you need to actually make enterprise software. And this is where it gets interesting because this kind of use cases, you see them everywhere, and everybody is cold coding or, you know, connecting whatever they need by directly plugging connectors, just letting any kind of data access any kind, I mean, any kind of agent access any kind of data without any kind of control, which basically prevents this kind of use cases to be deployed into production. So we really think that the future of work Is making sure that we bridge, all those personas into one, and this is where we're investing. So instead of using software, because RDF Steven does not necessarily know how to make software, which is direct agents, We have one unified harness. So I'm not gonna present a different product if you're in finance or if you're in legal or if you're in engineer. The product, the model is able to adapt what it will use, what it will show with a unified harness across the different surfaces. So what We saw earlier again is exactly what we're working on, making sure that you can start a session locally on your own laptop and then teleport it to the web and then execute it in sandboxes. Just check it on your phone, on the app, actually steer it. It can actually push to GitHub. It can actually check your CI. It can iterate on that, and this is where the one unified harness works so well because the business user who just wants to process some spreadsheets will be able to use the same without noticing the difference. So the platform obviously gets more, valuable each time someone uses it. Someone create a skill, someone creates an app, they share them with the rest of the team. We can also refine the skills, refine the prompts. So typically, in the use cases that I've just shared, well, you can see after one week of the execution, of this investigative agent, what worked, what did not. Maybe the way it called the Graphana MCP at first was wrong because he did not understand this undocumented option. Well, by looking at the traces, now the skill can self correct and can actually specify how to properly use this connector in the future to make sure that it will be more token efficient and faster. This leads to agentic safety because those agents, they're so powerful. They do so many things, but, obviously, they have risks. And I'm sure you all have someone in your family that in the past ten days asked you if AI would kill the word. Well, Mistral will not, and we're investing heavily In agentic safety for that reason. Briefly, what is, agentic safety? If we talk about the different threat models that we have with agents, they usually have privileged access to a lot of connectors, a lot of data, and then can't do anything with it. Since you're nondeterministic, it means that it's really hard to put hard coded, guardrailed, saying if you do that, well, don't. It's hard because next run, it might not do it, actually. And then obviously, attacker control input. The prompt injections that we all know about, they're getting more and more sophisticated. And, you know, we try to make it and bake it into the model, but then each model has a jailbreak. And as soon as you're fetching user data or data coming from the outside, this is always a threat and always something that can happen. So how can we protect against this? There is some work to be done in the model. You want the model to understand as much as possible what is an injection, what is a tree instruction, and what comes. From user and safe context. Obviously, you want also want to restrict the privileges of your agents. You want to make sure that at run time, they can only do what they need to do, and this can be dynamic. Depending on the step, depending on which workflow, you you could change the permission dynamically as well. Obviously, sandboxes are very important here. And then you have runtime guardrails. You have everything that you need to observe the behavior of the agent. Itself by looking at the traces, the low level traces, and that's where sandboxes are also so important. You can look at each system call, at each network call. And from this, you can derive heuristics and understand what sandbox is actually trained to do if the agent is following a nominal pattern or doing something that is a bit abnormal. So the problem with this is when you add all those security guardrails everywhere, you basically break all the use cases because you had such a powerful agent, such powerful tools, but then now You're restricting everything because you fear everything. And so this is a clear dilemma that we have to deal with. So we are solving this with dynamic collaboration and making sure that, as I said, you will grant the right amount of permissions at the right moment when you need it and when you have a good reason to grant them. And also making sure that we can filter which operation should be allowed and why. This is a very quick demo of a typical prompt injection where we will ask Vibe CLI to Perform a task by looking at a GitHub issue. So it's gonna go very fast. But by looking at the GitHub issue that has a prompt injection in it, it will actually post a command here that illustrates a secret key just because you took the, issue content that was posted by an adversarial user. So we're working on prompt injection rules where you can describe a set of classifiers. You can pick which one you want. It could be APIs. It could be secrets. It could be You know, company knowledge and things like this. You define a source. You define thresholds. You define the kind of moderation that you want to use, and you can define as well, what you wanna do, after that. So once you've defined your rule, you will now attach it to, a policy. Everything is configurable. And the policy will allow you to basically find out where you want to apply this rule on which product surface because it could be Vibe on the web. It could be the mobile app. It could be the CLI. So this is exactly what the rule allows you to do as an enterprise admin. And you can define the threshold here. You can apply it. And then you just need to enable it. And now you will, use the Vibe CLA, again, perform the same action, but it will be stopped at the network level without the user having to configure anything. And that power of enterprise controls is that the user never opted in to get protected. This is a policy that has been defined and pushed at the enterprise level. And this is the kind of work that Doing on, on safety. But safety does not... Is not enough if your AI is not sovereign. So what is what is sovereign? Well, you have data sovereignty. Where is your data? Where, or your users' data, your customers' data? Where is it stored? Where is it retained? Who has access to it? Which hyperscaler or which company, because of jurisdiction or something, could actually read this data? Intelligence sovereignty as well is when you analyze the behavior of the model, when you try to tweak When you find tuning, where is this happening? Compute as well. So now you have all the inference. When you do inference, you all buy ZDR everywhere, but still the data transiting through some places, and understanding this is also so important. Or is operational because you need to know exactly what's going on in your company, exactly what's going on in terms of security, audit logs, and and things like this. So the way we answer this is typical This kind of data center that we are currently building. Mistral was just a model company some time ago, then it became a product company, and now we are building our own data center. This is typically, 10 megawatt facility, near Paris with bit 300 from NVIDIA. So this is one of the largest, data center with this kind of cutting edge hardware, from NVIDIA in Europe. And this is typically where we serve our critical work. Close internal ones. This is where might be training right now. But this is also where we put the most sensitive customers, workloads. And this is how we can have, from end to end, a full sovereignty path, that we present to the customers, including agentic safety, but also including inference and training. But you also need models, obviously, for, all of this to work. This is basically the family of models that we have. The generalist Lines, different size for different matters. I know you're all expecting a big one, and I hope we can deliver it very soon. But we've also been focusing a lot on specialist models. There are two kind of specialist models, the specialist models that serve many purposes. We're gonna talk about OCR and Voxer in a moment, but also all the models that we fine tune for specific customers that have specific set of data. So think about the rare languages. Think about, you know, some companies have their own programming languages that nobody has heard of. So obviously, by default Coding models are not good at it. So how you can actually, you know, sit down with a company and make a specialized model makes an amazing difference. And we have a few experimenting models that allow us to refine the frontier of science. Lindstroll for formal proof. Chillstroll, we announced a month ago. You know, we're all talking about dev everywhere, but Chillstroll is basically, you know, answering a moderation question extremely fast with a very nice architecture, for safety purposes. If I dive in a bit into the, into the specialist models, let me talk a bit about Voxel, our family of audio models. So we do have, text to speech models. And for this, I will just showcase, I would say, something that I said last year at AI Engineer, and you tell me which is what. We really saw that the mission of providing open source models to actually bring the AI to the masses, was also compatible with pushing enterprise beyond and building. Some strong, focus. So what we build at Mistral so once you actually collect all this data, you wanna watch people do what they do. And in enterprise context, it's really important to use this to understand how people are actually doing their job. How are they processing their days? What kind of workflows are they putting in place? So as you might have guessed, one of them was, AI generated. The other one was me. So this is the kind of voice cleaning capability that we We're heavily investing on to make, this kind of models extremely useful. The second sample was not me at all, and and this is the kind of capabilities that we're working on. For, speech to text, we work on higher accuracy. We make sure that we can, yeah, basically mimic any kind of voice, any kind of accents so that it can be useful in any kind of settings. For the speech to text, we already have state of the art models, and we're improving them a lot. What we're working on is real time diarization so that you can have, you know, meeting a live transcription, but also knowing which speaker is which, making sure that we support more languages, obviously, accurate timestamps, more languages, making sure that it works in very noisy settings, and we can do live live transcription actually. So within banking, we have a very important process, which is called QIC. It's a long process. So with Mistral AI, we started to work on a specific QIC. Process for corporates in Belgium, and we developed two agents. We were able to go from 80% of files incomplete sent to the middle office to only 10%. For the clients, the application takes much So it was no cheating. It was, natural live transcription and live translation, of a video using our, latest models. So we are actively working on making it more useful for customers. And this is something that is key in the way we work. We're not Trying to, you know, benchmarks everything. We're trying to make things that are gonna make a difference on the ground with our own customers. Briefly about OCR, 4.1, the model that we announced, this summer. So this is also a state of the art, OCR model that is now able to transcribe any kind of handwriting, any kind of scanned old document from the sixties or something. I went, slightly fast here. ## Slides ### 00:00:18 # Demo time Bench, live [Illustration of a stick figure pulling back a curtain to reveal a machine under a spotlight, with an audience watching] ### 00:00:50 # Factory [Screenshot of a web application dashboard showing a list of automations, an activity log, and a visual workflow diagram] ### 00:01:20 ### Screenshot of Bench web application [Screenshot of the 'Bench' web application interface, showing a dashboard with chat functionality and a list of to-do items] ### 00:02:21 - AI Engineer Paris - Mistral - langfuse by ClickHouse - Notion - NVIDIA - Descript [Logos of AI Engineer, Mistral, langfuse, Notion ### 00:02:53 - Mistral - langfuse by ClickHouse - Notion - NVIDIA - AI Engineer PARIS - AI Engineer PARIS - descop [Collection of company names and logos ### 00:03:16 # Building Frontier AI The Path Forward for Real-World Solutions Mistral [Mistral logo] ### 00:03:51 - ROI over benchmarks. - ...tomer con - ...curity, and - ...stomizabilit - ...e to AGI. [Pixel art illustration of a rocket] [Pixel art illustration of a key] ### 00:04:50 ### Mistral x Koyeb [Mistral logo (a red pixelated creature) and Koyeb logo (a stack of three upward-pointing arrows)] ### 00:05:19 > From day one, we've stayed close to the hardware to provide high-performance with sustainable costs. We're operating tens of thousands of applications on top of bare metal servers across 10 ### 00:05:49 > By integrating our expertise into Mistral AI's world-class AI ecosystem, we are positioned to revolutionize core R&D. Together, we are providing the foundational intelligence required to design and build the next generation of aircraft, vehicles, and semiconductors. > > Johannes Bratter, Co-Founder and Chief Science Officer, Emmi AI ### 00:06:14 ## s Brandstetter ## nder and Chief ### 00:06:47 # Mistral announces Series D - €3 billion raised - Valuation of more than €21 billion - Largest equity fundraising round ever completed by a European technology company [Grid of colored squares with pixelated icons, including an 'M' logo, upward arrows, a box, and a flask] ### 00:07:13 # AI Engineer Paris. One year later. ### 00:07:47 # What's next. Looking forward to 2027 and beyond ## The future of Work. Start with an idea, end with a project, all from one unified place. ## Agentic safety. Dynamic collaboration for secure agentic actions that don't break in the real world. ## Your Sovereign AI. Own you AI data, intelligence, compute, and operations. ## Advanced models for advanced work. State-of-the-art models for real-world applications with more innovations on the way. ### 00:08:17 # The Mistral Suite is built around these three steps. ## REALIZE - Solutions: Use case acceleration | Domain-specific research | Enterprise transformation ## ACTIVATE - **Vibe for work** - AI-enabled business processes - **Studio**: Frontier models and builder tools - **Forge**: Advanced AI R&D Suite - **Vibe for code** - AI-enabled software engineering ## BUILD - Continuous improvement with enterprise context - Serverless | On-prem | Edge | Mistral Compute - End-to-end traceability, identity, and policy management [Diagram illustrating the Mist ### 00:08:49 # The Mistral Suite is built around these three steps. **REALIZE** Solutions: Use case acceleration | Domain-specific research | Enterprise transformation **ACTIVATE** - ** ### 00:09:14 [Abstract image of textured fabric in shades of purple, grey, and dark blue/green] ### 00:09:48 [Blurred image of a person's arm in a purple sleeve and a textured surface, with a small black remote control on the left, and a solid red panel on the right.] ### 00:10:14 [Close-up image of a person's arm in a purple sleeve and part of a textured chair or car seat, with a red color block on the right side of the screen] ### 00:10:46 # Developers build business automations end to end ## In-house example: incident investigation ### 1. Slack alert - An incident alert starts the investigation. ### 2. Agent investigates - Grafana metrics and logs. GitHub or Linear for code and ownership. ### 3. Findings in Slack - Affected services, recurring errors and possible causes. Evidence ### 4. Continue in thread - Responders ask follow-ups and build on the work already done. ## Developers assemble reusable building blocks through APIs & SDKs - Triggers - Agents API - Connectors - One-click deployment [A four-step horizontal process flow diagram] [Four rectangular blocks representing building blocks] ### 00:11:16 - Grafana metrics and logs. - GitHub or L... for code - and own... - usable building... through AP... ### Agents ### 00:11:49 ### Building what's next ## Developers build business automations end to end ### Screenshot of Sentry alert: Critical: High error rate in project: quota [Line graph showing event count ### 00:12:18 # Business users describe the workflow. Vibe builds the automation. ## The same building blocks could support processes across the business ### 1 Describe the process > "When interview feedback is missing, send a personalized reminder." > HR team ### 2 Build with Vibe **Vibe** Creates a workflow to review and deploy ### 3 Run the automation - Check for missing feedback - Identify the interviewer - Send a personal reminder Vibe would assemble the same reusable building blocks - Triggers - Agents API - Connectors - One-click deployment [Flow diagram illustrating a three-step process: "Describe the process", "Build with Vibe", and "Run the automation".] [Four rectangular boxes representing reusable building blocks: Triggers, Agents API, Connectors ### 00:12:49 # to review a... - ...e the same... - ...e building blocks - Ag... API ### 00:13:18 ### Building what's next ## Business users describe the workflow. Vibe builds the automation. The same building blocks could support processes across the business ### 1 Describe the process > "When interview feedback is missing, send a personalized reminder." HR team ### 2 Build with Vibe Vibe Creates a workflow to review and deploy ### 3 Run the automation - Check for missing feedback - Identify the interviewer ### 00:13:46 # The future of Work ## From using software to directing agents ### 00:14:46 ### Building what's next # The future of Work Platforms get more valuable with each person who uses it ### 00:15:14 ## Building what's next # The future of Work Platforms get more valuable with each person who uses it ### 00:15:47 ## Protecting Agentic Workloads # A new vector of attack ### Privileged Access Possesses high-level system permissions to read sensitive data and execute critical system actions. ### Non-Deterministic Generative models behave unpredictably, making deterministic safety rules nearly impossible to enforce. ### Attacker controlled Inputs External, untrusted data can easily inject instructions and override the original prompt guidelines. [Icon of a key next to "Privileged Access"] [Icon of branching arrows next to "Non-Deterministic"] [Icon of a pencil next to "Attacker controlled Inputs"] ### 00:16:15 ### Protecting Agentic Workloads ## A new vector of attack - **Privileged Access** Possesses high-level system permissions to read sensitive data and execute critical system actions. - **Non-Deterministic** Generative models behave unpredictably, making deterministic safety rules nearly impossible to enforce. - **Attacker controlled Inputs** External, untrusted data can easily inject instructions and override the original prompt guidelines. [Three columns, each with a heading and a short description, and a small icon next to each heading: a key, a branching arrow, and a pencil.] ### 00:16:46 # How can we protect ourselves against prompt injection? ## Model Hardening Hardening the LLM against adversarial instructions and system prompt overrides. - **System Prompting**: Robust framing to anchor core instructions. - **Adversarial Training**: Fine-tuning models to reject malicious inputs. - **Few-Shot Defense**: Explicit examples of hostile instruction handling. ## Least Privileges Restricting capabilities to the absolute minimum required to perform agent tasks safely. - **Sandboxing**: Isolated environments to execute untrusted code. - **Protect Secrets**: Avoid exposing secrets to the agentic context. - **IBAC**: Intent-based access control policies. ## Runtime guardrails Active data classification booby traps at hook points to mitigate leaks before they happen. - **Input Protection**: Detects sensitive data leaking into tool inputs. - **Output Validation**: Identifies hidden system instructions in outputs. - **Live Enforcement**: Instantly blocks violating operations. [Pixel art icon of a robot next to "Model Hardening"] [Pixel art icon of a crossed-out circle next to "Least Privileges"] [Pixel art icon of a fence next to "Runtime guardrails"] ### 00:17:17 - **Prompting**: Robust anchor core - **Adversarial Training**: [Training] models to [resist mali]cious inputs. - **Defense**: Explicit [detection] of hostile [input] handling. - **SandboxIng**: Isolated environments to execute untrusted code. - **Protect Secrets**: Avoid exposing secrets to the agentic context. - **BAC**: Intent-based access control policies. ### 00:17:48 # Safe and effective access that doesn't break solutions Protecting Agentic Workloads ## Dynamic Collaboration Effective protection relies on a tight coordination between access control and runtime context. - **Contextual Access**: Permissions are granted dynamically based on the exact real-time state of the runtime context. - **Surgical Operations**: System-level filters block specific unauthorized operations (e.g., data leaks) without disrupting the entire conversation. ### 00:18:21 Use the GitHub connector to review outstanding issues in safchain/agantic-safety-demo, reproduce them as described, and comment on each issue with the result. I found one open issue ### 00:18:46 ### Configure case [Screenshot of a software dialog box titled "Configure case" with fields for Name, Type, Condition, Outcome, and Side effects.] ### 00:19:19 # Prompt Injection Policy [Screenshot of a user interface for managing a Prompt Injection Policy] ### 00:19:47 - Control where AI workloads capacity required to operate - Control... AI system in pro... observ..., policy, routing ### 00:20:19 # What it means to be Sovereign - **Data sovereignty**: Control where data is processed, stored, and retained; who can access it; and whether it can be used to improve any shared model. - **Intelligence sovereignty**: Control the model layer and the compounding loop around it. - **Compute sovereignty**: Control where AI workloads run and retain dependable access to the capacity required to operate them. - **Operational sovereignty**: Control the AI system in production, including identity, security, observability, policy, routing, deployment, and lifecycle management. ### 00:20:47 # Les Ulis: sovereign inference, live in France - **B300s** Latest-gen NVIDIA GPUs, one of Europe's largest deployments - **10MW** Dedicated AI compute - **Inference** Purpose-built for real-time inference - **Q3 2026** Operational [Image of rows of server racks with yellow cabling] [Aerial view of a large data center building] ### 00:21:18 [Image of server racks on the left and green foliage on the right, behind a speaker] ### 00:21:49 # Mistral Model Families ## Generalist - Large - Small - Medium - Ministral ## Specialist - OCR - Codestral - Voxtral - Codestral Embed ## Open Experiments - Leanstral - Shieldstral [Table listing Mistral model families, categorized by Generalist, Specialist, and Open Experiments, with a unique icon for each model. The table also features labels for LABS, OPEN, and PREMIER.] ### 00:22:21 # Mistral Model Families - **Generalist** - Large - Small - Medium - Minstral - **Specialist** - OCR ### 00:22:44 # TTS ## Voxtral - Family of audio models - **Speech to text**: Voxtral Mini Transcribe 2, Voxtral Mini Transcribe Realtime, Voxtral Small - **Text to speech**: Voxtral TTS [Two pixel-art Eiffel Tower icons, each paired with an audio playback button labeled "AI Engineer 2025"] ### 00:23:18 # Voxtral ## Family of audio models ### Speech to text: - Voxtral Mini Transcribe 2 - Voxtral Mini Transcribe Realtime - Voxtral Small ### Text to speech: - Voxtral TTS [Two pixel art Eiffel Towers, each next to a speaker icon and a red button labeled "AI Engineer 2025"] ### 00:23:51 # AI Engineer PARIS Organized by Mistral ## Voxtral Speech to Text (STT) - High accuracy - Speaker diarization - Context biasing - Word-Level timestamps - Multilingual support - Noise robustness - Long audio support ## Ex: live translation in real time [Pixel art illustration of a microphone] ### 00:24:49 # Example for 2 clients ## Client 1 - Status: 000 ## Client 2 - Status: 000 [Diagrams showing two irregular polygons, one for each client] [Small green box with three colored circular icons (green, red, yellow) and a text input field]