Cisco AI

What if the future of AI isn't about building bigger models, but smarter, more specialized ones? In the latest episode of Cisco Champion Radio, Cisco Foundation AI researchers Supriti Vijay and Aman Priyanshu take us behind the research of Antares, Cisco's purpose-built AI model for vulnerability localization. Hear directly from the researchers as they explore: ➡ Specialization over scale: How smaller, purpose-built models can rival or outperform much larger frontier models on specialized tasks. ➡ The engineering behind Antares: From curated training data and reinforcement learning to autonomous tool use. ➡ Rethinking AI evaluation: Why the team developed VLocBench to better benchmark real-world vulnerability localization. ➡ Lessons learned: The breakthroughs, challenges, and experiments that shaped Antares. 🎙️ Go behind the research and hear what's next for specialized AI in cybersecurity: https://smarturl.it/S13E14

We are releasing the weights of a new and improved Mistral Large 4, advancing the open frontier in coding and cybersecurity. Trained from scratch with clear lineage in European data centers, Le Chonk provides frontier-level performance in enterprise-critical domains, and gives organizations a trusted foundation for Sovereign AI. 🛡Cybersecurity: Enterprise and public sector institutions can now download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure. The models can run on a company’s own infrastructure and no third party can restrict access or disable them. Industry-leading cybersecurity companies, including Cloudflare, CrowdStrike, Zscaler, Cisco, Elastic, Fortinet, Rubrik and Orange Cyberdefense, are supporting and testing ML4. 💻 Coding: ML4 delivers frontier coding performance with significantly fewer tokens. It leads Cognition’s FrontierCode 2.0, which measures whether a maintainer would merge the agent’s change, and matches the best open-weight models on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. 💻 Work: ML4 also keeps improving on everyday enterprise tasks, with 73.3% on Finch for finance spreadsheet workflows and 84.5% on BrowseComp for web browsing. Mistral Large 4 will serve as the foundation for a new generation of specialized and open Mistral models. We’re committed to giving developers, enterprises, and institutions the power to own, customize, and control their AI.

When there’s no shortage of AI agents and tools to choose from, how do you know which ones are safe to put to work? We’re launching the ServiceNow Control Tower Alliance with leading AI, cloud, and security companies to make finding solutions you can trust simple. Together, we will build AI solutions on the Super Intelligence Control Tower, with shared standards for agent identity and safety guardrails. For customers, that means a curated set of solutions, qualified and certified by ServiceNow before they’re listed, so you can adopt AI with governance in place from the start. Founding members of the Control Tower Alliance include: ➡️ Anthropic ➡️ Amazon Web Services (AWS) ➡️ Cisco ➡️ HiddenLayer ➡️ Kong ➡️ NVIDIA ➡️ Okta ➡️ OpenAI ➡️ Stripe Partner Ecosystem ➡️ WitnessAI ➡️ Zenity Learn more: https://lnkd.in/eXefgw2J

Securing AI at scale takes more than visibility. It takes connected security and governance. Cisco is joining the ServiceNow Control Tower Alliance as a founding member, helping enterprises securely adopt and govern AI at scale. Together, Cisco AI Defense and ServiceNow connect AI security insights with enterprise governance and remediation workflows. Here's what that means for customers: ➡ AI Discovery: Identify models, agents, datasets, and MCP servers across cloud environments. ➡ Proactive Security: Scan AI components for risks and test models and applications for vulnerabilities. ➡ Runtime Protection: Apply security, privacy, and safety guardrails to AI applications and agents. ➡ Connected Governance: Bring AI security findings into ServiceNow workflows to assign ownership and track remediation. And this isn't just a future vision. Two Cisco AI Defense integrations are already available in the ServiceNow Store, connecting security insights with actionable governance workflows today. Learn how Cisco and ServiceNow are helping enterprises scale AI with confidence: https://lnkd.in/gDkeAksS

AI agents chat with you, and they also read, reason, and act on your behalf. As agent conversations grow longer, security threats may unfold across multiple steps that traditional defenses focused on a single point in time may miss. Introducing Project Quarry, a new open source detection language from Cisco designed to help security teams identify suspicious patterns across AI agent activity. ➡ Detect multi-step threats: Identify risky behavior across conversations, tool calls, and agent workflows. ➡ Combine intelligence with precision: Bring together AI-powered detection and rule-based checks to identify potential threats. ➡ Turn findings into actionable evidence: Give security teams visibility into what happened, why it matters, and which actions triggered a finding. Already in production within Cisco AI Defense, Quarry is designed to provide a shared foundation for agent security that the broader community can build on. Understand the context in which agentic threats may emerge. Learn more about Project Quarry and explore the open source release: https://lnkd.in/gXH-bhpG

The next leap in AI isn't just about choosing the best model. It's about applying the right one.    Introducing Cisco Model Foundry, a new platform service within Cisco Cloud Control that brings Cisco's purpose-built AI models to our partners.    With Model Foundry, partners can:  ✦ Access Cisco's domain models, built on decades of networking, security, and operations expertise.  ✦ Access models deployed in Cisco Cloud Control and integrate inference capabilities into customer applications, agents, and workflows through OpenAI-compatible APIs. ✦ Customize models with customer-specific data through fine-tuning, with built-in evaluation to measure results.    And we're already seeing the potential. In lab testing with World Wide Technology (WWT), a fine-tuned Cisco Deep Network Model outperformed a frontier model by 25% on real-world troubleshooting scenarios, at a fraction of the cost.    Announced today at Cisco Partner Summit, Model Foundry is now in Controlled Availability, starting with partners.    Read the latest from DJ Sampath on how we're bringing purpose-built AI models to our partners: https://lnkd.in/gPi-bWfs

How do you evaluate AI agents that take multiple steps, use tools, and make decisions along the way? Introducing Agentic FAITH, an extension of Cisco Foundation AI's FAITH evaluation framework that brings question-answering and agentic cybersecurity evaluations into one workflow. With Agentic FAITH, researchers and engineers can: 🔹 Run evaluations through one framework: Use the same workflow for question-answering benchmarks and multi-step agentic tasks. 🔹 Evaluate vulnerability localization: Use VLoc to assess how well agents locate vulnerable files and avoid false alarms on patched code. 🔹 Analyze results: Inspect tool calls, execution records, and outcomes to better understand agent behavior and performance. Read the latest from Cisco Foundation AI Researcher, Kimia Majd, to explore Agentic FAITH and get started: https://lnkd.in/gX5rsWyk

As local AI shifts from chat to action, enterprises need more than powerful compute. They need visibility, security, and control across their AI workloads. Cisco and NVIDIA are bringing together Cisco Cloud Control and NVIDIA DGX Station for Windows to help organizations govern local AI agents, monitor their behavior, and enforce security policies across their AI fleet. Here's how the solution comes together: - Agent observability: Splunk Agent Observability provides visibility into agent activity, tool calls, token usage, and potential security risks. - AI security and enforcement: Cisco AI Defense and DefenseClaw help apply security guardrails and enforce policies around agent actions. - Centralized governance: Cisco Cloud Control brings agent behavior, policy state, and AI workload insights into a common interface, helping teams investigate risks and apply policies across their AI fleet. From individual workstations to enterprise-wide AI infrastructure, Cisco and NVIDIA are working together to help organizations build resilient, secure AI environments. Learn how Cisco Cloud Control and NVIDIA DGX Station are enabling local AI resilience: https://lnkd.in/ghszxRwe

Did you know there's actually a leaderboard that ranks AI models based on how secure they are? 👀 Because not all AI models are created equal. And some are much easier to manipulate than others. Cisco has developed an LLM Security Leaderboard that tests and compares over 120 large language models to see how well they stand up against threats like prompt injection attacks and jailbreaks. And I got to sit down with Amy Chang, Cisco's Head of AI Threat Intelligence and Security Research to learn how it works. We also talked about a few things I think everyone working with AI should understand: 🔹 AI Security Taxonomy: Cisco has mapped out 19 objectives, 40+ techniques, and 110+ sub techniques to help organizations understand how AI systems can be attacked. 🔹 Human oversight: Why keeping humans in the loop matters, especially as AI agents start making decisions and taking actions for us. 🔹 AI Bill of Materials: Why organizations need to know where their AI models came from, how they were trained, and what they're connected to. One of my biggest takeaways from this conversation? Just because AI CAN do something doesn't mean we should automatically give it permission to. And if you're using AI in your business, understanding the security of the models you're using is a pretty good place to start. 🔗 Check out Cisco's LLM Security Leaderboard: https://lnkd.in/ed8E_e-H 🔗 Explore Cisco's AI Security Taxonomy: https://lnkd.in/ezvJACes I'm curious: Have you ever thought about how secure the AI model you're using actually is? #AISecurity #Cybersecurity #ArtificialIntelligence #Cisco #LLMSecurity Cisco AI, Cisco Security

Agents need access to do their job. They don't need access to everything else. Cisco's Secure AI Factory with NVIDIA just expanded to rack scale through a partnership with Supermicro, building s#ecurity directly into AI infrastructure. Learn how DPU-based enforcement keeps agent permissions scoped and auditable from Cisco SVPs Will Eatherton and Jeremy Foster ➡️ https://cs.co/6043BHVn4L #GTCBerlin