Hi, everyone, in this video we will explore Cisco's AI defense. AI defense is a comprehensive purpose built AI security solution that mitigates the risks associated with the development and deployment of AI models, applications, and agents.00:05At its core, AI defense empowers your security teams to discover AI assets across distributed cloud environments, detect vulnerabilities through supply chain scanning and algorithmic testing, and protect runtime applications with security, privacy, and safety guardrails updated with the latest threat intelligence. AI defense extends protection to a gentic AI system.00:20And the model context protocol MCP. This includes scanning MCP servers for compromised or malicious assets, enforcing runtime protections across MCP requests and responses.00:40With the rise of AI, new security threats emerge. AI security is different from traditional cybersecurity because AI itself is different from traditional software. There are several differences, but these four are key. AI is non deterministic, unlike traditional software, where a single input is programmed to yield.00:52Consistent outputs, AI has an element of unpredictability.01:12AI is adaptive. Traditional software doesn't change without new code being pushed live. AI can change for several reasons, including in context learning, fine tuning of a model, and changes to connected data sources. This means new behavior and vulnerabilities can emerge unexpectedly. AI is data centric. Traditional software has a distinct separation between the.01:16Code and data layers. With AI, a model's intelligence comes from the data used to train it, meaning these distinct lines NO longer exist. While traditional software vulnerabilities can be patched, AI vulnerabilities are far less straightforward to fix. AI is autonomous, more specifically, agentic systems are capable of unprecedented levels of planning.01:36Reasoning and acting with little to NO human supervision. This offers tremendous productivity advantages, but also massive security challenges.01:56AI applications are fundamentally different than traditional software applications. Generative models and agents introduce non deterministic elements into the application stack, meaning the same input can result in different outputs. This creates new safety and security risks vectors that aren't addressed by existing cyber security solutions such as.02:06Unsafe or inappropriate responses which can have outsized consequences, data exfiltration or leakage given retraining, fine tuning and connected data sources, downloading model files from HuggingFace without proper inspection, not protecting against cyber attacks.02:25Failing to comply with AI security standards and regulations. If left unchecked, these AI risks can lead to direct negative impacts on your business. AI defenses framework for AI security has three steps. Discover what AI assets exist in your business. Detect their risks and.02:40And protect them with guardrails that cover these gaps.03:01AI defense provides protection across the AI lifecycle. AI inventory provides an inventory of all AI assets, inventory, the AI models, agents, and connected data sources across distributed environment to understand their usage and risk, AI supply chain risk management scans, open source components for security threats, scan model files.03:06Repos and MCP servers to proactively block malicious or unsafe AI assets before operations are impacted. AI model and application red teaming to test for vulnerabilities, identify safety and security vulnerabilities across models at scale with algorithmic red teaming technology. AI runtime protection secures applications in production.03:26Protect production all apps and agents with guardrails embedded in the network, block attacks and harmful responses in real time.03:46By integrating AI defense within their workflow, AI dev teams will seamlessly extend security into all stages of the development life cycle. This promotes a shift left approach to AI security, incorporating important measures like supply chain scanning and model red teaming earlier on in AI development.03:55Now let's explore AI supply chain risk management. Supply chain begins with discovery, cloud service providers discovery of AI workloads in AWS, Azure, and Google Cloud, AI bill of materials scanning and containers for embedded AI artifacts connectors to MCP and model repositories visible.04:13Ability is the foundation of defense.04:33AI bill of materials results from the agenic discovery of AI assets across your code base, uses llmiss to traverse source code, identify AI assets like agents, MCP servers, tools, and skills, and map how they interconnect. Discovered assets feed into AI Defenses enterprise inventory, where the platform correlates relationship.04:37Across your AI landscape.04:57Runs locally in your build pipeline with zero cloud tethering. The AI defense CLI generates the AI bomb on premise and results are uploaded to AI defense enterprise SaaS only after your team approves. AI supply chain risk management scans 3rd party assets like open source models to identify vulnerability.04:59This ensures you can build AI applications with components that aren't compromised, automatically scan model files in your private repositories to identify vulnerabilities like code execution and suspicious imports, scan MCP servers to inventory tools and detect tool poisoning attacks and prevent the usage of insecure models and 3rd party assets.05:19AI inventory provides a foundational understanding of what AI assets exist within your cloud environment and externally is important usage context around them. Automatically uncover AI assets across your distributed cloud environment, including models, agents, and connected data sources.05:40Next, you want to identify vulnerabilities in your AI models and applications. With algorithmic red teaming technology, our product accomplishes in minutes what would otherwise take weeks of manual testing. AI Defense's risk focused red teaming process allows advanced multi turn techniques.05:58Which simulate realistic conversational attack scenarios that evolve over multiple turns, allowing you to test your AI applications resilience against sophisticated adaptive threats that go beyond single prompt interactions.06:18AI red teaming can be performed in two different modes, standard and adaptive. Standard to evaluate system defenses against libraries of known attacks or adaptive to simulate adversarial threats to expose application specific attack goals for your unique business risks and target endpoints.06:32Automatically evaluate models and applications for 200 plus security and safety combinations. Enhanced security posture with a risk centric interface that delivers granular technique breakdowns, automated risk segmentation, and quantified performance metrics mapped to industry standards to prioritize mitigation, available for multiple languages and three different modalities.06:50Please.07:10Now you know what AI assets exist in your environment and what they're vulnerable to. AI runtime protection is the final piece, providing guardrails that cover your models against the risks you previously identified. Define bi directional guardrails for applications and agents that block malicious prompts and unsafe responses, configure guardrails to.07:12Cover specific model vulnerabilities and fit unique AI applications stay protected against rapidly evolving AI threats, including those to MCP servers.07:32AI runtime protection provides guardrails with broad coverage that fall under three main categories. Security guardrails block attempts to override an llm's alignment, generate or retrieve code, maximize resource consumption or execute other insecure outcomes.07:43Privacy guardrails block attempts to reveal sensitive information that might cause harm if leaked to an unauthorized party. Safety guardrails prevent the generation of content which may offend or cause harm to individuals, businesses, and broader society. Static guardrails defend against universal threats. Adaptive guardrails hand.07:59Little risks unique to each application or agent. Adaptive guardrails can be built, tested, and refined via policy studio.08:19AI defense guardrails be configured to fit your unique business and AI applications. Build and enforce context dependent business rules essential for compliance, brand identity or operational scope. AI runtime protection helps protect agenic and MCP connections too. In addition to applying our existing guardrails to agents.08:28Interactions. We have guardrails purpose built to mitigate agent and MCP specific threats like tool poisoning.08:48AI defense can be deployed in multiple setups to cover every situation. AI Defense offers the convenience of a Cisco hosted cloud managed control plane with the flexibility of deployment options for the data plane. SAS VPC or AI pods. SAS for customers looking for a simple, flexible deployment with.08:56Zero infrastructure to manage.09:16VPC for customers looking to balance data control and compliance with cloud scalability, and the customers that wanna manage AI workloads themselves rather than relying on hyperscalers will be able to do it on their own data center. By integrating Cisco AI defenses discover, detect and protect use cases into your AI strategy, you equip your organization to push.09:18The boundaries of innovation safely, swiftly, and with complete confidence in your security risk posture. Thank you for watching and make sure you check the other AI defense cyber capsules.09:38
