Cisco Enhances AI Security with Claude Mythos Red Teaming and Advanced Agent Observability

Cisco has unveiled new AI security and monitoring capabilities, announced on November 9th in the US and coinciding with Cisco Live APJC in Melbourne on November 10th. A key highlight is the integration of Anthropic's Claude Mythos into Cisco's AI Defense service. This will enable authorized enterprises to conduct structured and auditable red-teaming exercises on their own AI agents. Additionally, Splunk Agent Observability is receiving new tools designed to correlate token spending with evaluation quality, including specialized evaluation models and tuning functionalities. This development is crucial for organizations deploying AI agents, as it directly addresses the escalating concerns around AI security and operational transparency. As AI systems become more autonomous and integrated into critical business processes, the ability to proactively test for vulnerabilities and understand their real-world performance metrics becomes paramount. Security teams and MLOps engineers are directly affected, gaining tools to harden their AI deployments and optimize resource utilization. The focus on auditable red-teaming also helps meet increasing regulatory and compliance demands for AI systems. This move by Cisco aligns with a broader industry trend towards more robust MLOps practices, particularly in the context of generative AI and AI agents. The industry is moving beyond basic model deployment to focus on the entire lifecycle, including continuous evaluation, security, and governance. The need for policy-as-code and automated model governance is becoming increasingly evident as manual oversight becomes unfeasible with the scale of AI deployments. Furthermore, the emphasis on connecting operational costs (like token spending) with performance quality reflects the growing maturity of MLOps, where ROI and efficiency are critical considerations. In practice, MLOps practitioners should leverage these new capabilities to establish more rigorous security pipelines for their AI agents. The red-teaming functionality offers a structured way to identify and mitigate potential adversarial attacks or unintended behaviors before they impact production. The enhanced observability tools will allow for more granular cost analysis and performance optimization, moving beyond simple API call counts to actual task-level success metrics. Teams should investigate how to integrate these new features into their existing CI/CD pipelines for AI, ensuring that security and performance evaluations are continuous and automated. This also underscores the importance of developing in-house expertise in AI security and agent evaluation methodologies.