Cisco plans to add Anthropic’s Claude Mythos to AI Defense and introduce Splunk tools for tracking AI-agent costs and output quality, giving security and IT teams more visibility as workplace AI use expands.
Cisco says Mythos will help AI Defense plan customized, multi-stage red-team exercises for customers’ AI applications and agents. AI Defense would manage the process: tests would be confined to systems registered by customers and within their authorization, with each run linked to an authenticated operator in an audit record. Customers would not get an open interface to prompt Mythos directly. Cisco plans to begin with select enterprises, then broaden access to qualified customers, though timing and scope may change.
AI Defense launched in January 2025. Cisco says the service has helped hundreds of customers identify flaws in AI models, applications and agents before production.
Cisco says Eval-G1 can reduce the cost of evaluation and guardrail runs by up to 98% compared with a commercial large language model. It also says Eval-G1 returns responses 2.5 times faster than traditional large language models. The Cisco announcement does not describe the benchmark methods behind those comparisons. The models and Evaluation Tuner are generally available starting in November 2026, Cisco says.
Cisco also plans to bring Cisco IQ vulnerability, lifecycle and configuration information into Splunk through APIs, allowing security teams to assess threat detections alongside data about potentially exposed infrastructure. Planned connectors would pass that context to supported AI agents and ServiceNow, where findings could be attached to problem records linked to affected assets.
Splunk’s 2026 survey of 2,050 IT operations and engineering professionals found that 81% of respondents said monitoring AI workloads had increased collaboration with security, while 55% said their organizations could not accurately control AI costs. Cisco did not disclose prices or measured customer savings for the new features.
Tokenomics will relate agent usage data from Amazon Bedrock and Amazon CloudWatch to output evaluations and business measures chosen by customers. Cisco plans a private preview in December 2026, followed by general availability through AWS Marketplace in early 2027.
