Welcome to this version of cybersecurity capsules on AI defense, where we will explore the programmatic security capabilities available through the Cisco AI defense Python SDK. Our agenda for today focuses on an introduction of the python.00:04SDK tool.00:20Followed by an overview. We will explore the supported platforms and the steps to install and configuring the SDK, and we will close with the main capabilities. This session will provide an in depth understanding of how to leverage the SDK for embedding AI security inspections directly into your Python applications, protecting LLM interactions.00:21AI agent tool calls.00:41In model deployments, let's begin this journey into enhancing our security posture through the power of developer integrated AI defense. In this slide, we are discussing the Cisco AI defense Python SDK. This software development kit is designed to simplify and accelerate the process of securing AI powered applications against emerging.00:42Threats such as prompt injection.01:02Data leakage and malicious model files. Its primary advantage lies in its developer friendly approach, which integrates seamlessly into existing Python applications with minimal code changes, in many cases, just two lines of code. The SDK supports inspection of chat conversations, HTTP requests and responses, model context protocol tools.01:04Calls and AI model files.01:24Importantly, it provides structured results with detailed classifications, facilitating easy integration into automated security workflows and CICD pipelines. For use an AI defense API token and endpoint URL are required. Let's discuss the supported platforms and providers for the AI defense.01:26Python SDK. The SDK requires Python 3.9 or newer and is compatible with all major operating systems.01:46Linux, Mac OS, and Windows. It is distributed via pypi, the standard Python package repository, making installation straightforward across all environments. The SDK supports auto patching for a broad range of LLM providers, including OpenAI, Azure OpenAI, Amazon bedrock, Google Vertex AI, coherent.01:53Mr. AI and light LLM. Additionally, it integrates with major agent frameworks such as line chain, line graph, Crew AI, autogen.02:13And the Open AI agents SDK. For managed runtimes, the SDK supports Amazon bedrock agent core, GCP Vertex AI agent engine, and Microsoft AI foundry. This versatility ensures that a broad range of developers can leverage the SDK regardless of their AI provider or framework of choice.02:23In this segment, we'll cover how to install the Cisco AI defense Python SDK. Begin by installing the package from pypi using PIP. Note that the pypi package name is Cisco A defense SDK, but you import it as AI defense in your Python code.02:43For local development and contribution, you can clone the repository from GitHub and install in editable mode. The project uses poetry for dependency management. After installation, you can verify the setup by importing the SDK in a Python session. This slide covers the critical aspect of authentication and configuration for the SDK.02:58Successful operation requires an AI defense API token and an endpoint URL pertinent to your environment. From a security standpoint, the SDK ensures that all inspection traffic is transmitted securely to Cisco AI defense for analysis. Note that NO credentials are embedded within the SDK itself.03:18Authentication is managed through tenant scoped API tokens. You have three methods to provide credentials. 1st, using environment variables, which is recommended for CICD workflows. Second, using a YAML configuration file such as agentsec.yaml.03:35Which provides comprehensive control over modes, gateways, retry policies, and timeouts. 3rd, directly via Python code when calling agentsec.product or initializing inspection clients. Each method offers flexibility depending on your operational environment and security practices.03:51Remember to use your specific region's base URL during these configurations. Here, we detail what the Cisco AI defense Python SDK accomplishes. This tool is designed to provide real time security inspection of AI interactions across multiple dimensions.04:09Agent runtime SDK, auto patches LLM clients with two lines of code for transparent inspection. Chat inspection, analyses prompts, responses, and conversations for security risks. HTTP inspection inspects HTTP requests and responses to AI APIs.04:26MCP inspection secures model context protocol tool calls in AI agents. Model scanning scans AI/ML model files for malicious code and threats. AI bomb generates and manages AI bill of materials for applications.04:44Management API, programmatic management of applications, policies, and events. The SDK operates in two modes. Monitor mode logs violations without blocking and enforce mode actively blocks requests that violate security policies. Results are returned as structured Python objects with detailed classifications, optimal for integration.04:59Integration into automated processing and reporting systems.05:19Now let's explore the agent runtime SDK, which is the recommended approach for most applications. The agentsec.product function automatically patches supported LLM clients so that all AI interactions are inspected transparently. The key requirement is to call product before importing your LLM client.05:24The examples demonstrate how you can enable protection with minimal effort. The SDK supports both API mode, which performs side channel inspection and gateway mode, which proxies all traffic through the AI defense gateway for server side enforcement. This flexibility allows teams to adopt the approach that best fits their security requirements and infrastructure.05:41The chat and HTTP inspection commands provide granular control over AI interaction security. With a chat inspection client, you can inspect individual prompts, responses or full conversations for a range of threats including prompt injection, PII, and PCI exposure.06:00Harassment hate speech, toxicity, and code detection. The HTTP inspection client extends these capabilities to raw HTTP requests and responses, supporting integration with the Python request library for seamless inspection of API traffic. Using these clients, you can build custom inspection workflows tailored to your application's specific sec.06:18Security requirements.06:38The MCP inspection capability focuses on securing model context protocol interactions, as AI agents increasingly use external tools and resources through MCP, they introduce new attack surfaces. The SDK inspects MCP JSON RPC 2.0 messages for security, privacy, and.06:40Safety violations in tool calls, resource access and responses.07:00Using the SDK, you can register and monitor MCP servers, secure tool call traffic through the gateway, and manage server security efficiently. These capabilities support continuous oversight of AI agent infrastructure ensuring that tool calling agents operate within defined security boundaries.07:04With a model scanning capability, you can scan AI and ML model files for security threats, malicious code, and vulnerabilities before deployment. With the AI bomb capability, you can manage the AI bill of materials for your AI and ML applications.07:22This functionality allows you to generate, submit, and manage AI bomb reports, which are crucial for maintaining an overview of all components within an AI or ML application. This command suite supports component listing and provides a summary of bomb statistics enabling comprehensive management of AI or ML projects.07:37Let's review the benefits the SDK provides for both security operations and the business. For security operations, real time visibility into all AI interactions across the organization policy enforcement at the application layer without infrastructure changes. Detection of prompt injection, data leaks, and toxic content centralized AI security posture management.07:57Through the management API complete audit trail of all inspected AI traffic for the business. Accelerate AI adoption with built in security guardrails reduce risk of data breaches through AI channels meet compliance requirements including GDPR.08:17HIPAA and PCI DSS open source SDK with NO vendor lock in on the client side works with any LLM provider, OpenAI, Azure, AWS, GCP, and more. To get started with the Cisco AI defense Python SDK, the following resources are available.08:31GitHub repository, Cisco i defense/idefense Python SDK, pipackage, pip install Cisco A defense SDK. Examples over 20 runnable examples covering all capabilities. Configuration guide, comprehensive YAML and programmatic configuration documentation.08:51The adoption path is straightforward. Install the SDK, enable monitor mode on a single application, review the inspection results, then scale to enforce mode across your AI portfolio. Thank you for taking the time to engage with our presentation on the Cisco AI defense Python SDK.09:09We encourage you to join the conversation in the comments below and share your thoughts or questions. See you on the next one.09:27
