New red-teaming, agent evaluation and infrastructure intelligence capabilities aim to help enterprises secure autonomous AI systems, control costs and accelerate vulnerability remediation.
As enterprises move from AI experimentation to autonomous agents operating across business applications and critical infrastructure, Cisco is expanding its security and observability portfolio to address the risks of agent-driven operations. Announced at Cisco Live APJC in Melbourne on November 10, the updates bring together frontier AI red teaming, agent observability and infrastructure exposure intelligence to help organisations govern AI deployments at scale.
A key development is the integration of Anthropic’s Claude Mythos into Cisco AI Defense. The model will augment the platform’s adaptive red-teaming framework with advanced reasoning to design contextual, multi-step attack scenarios against registered AI applications and agents. The objective is to uncover vulnerabilities before systems enter production.
The framework includes structured access, audit trails and repeatable testing. Cisco says engagements will be restricted to customer-owned and explicitly authorised targets, with each exercise tied to an authenticated enterprise operator. Testing sessions will have defined objectives, time limits and the ability to be halted on demand. The Claude Mythos capability is expected to become available soon to select enterprises.
“By uniting frontier-level AI runtime protections with machine-speed agent quality controls, we are building a continuous trust loop: protecting agents from compromise while ensuring every token spent delivers measurable business value.” — Jeetu Patel, President and Chief Product Officer, Cisco
Cisco is also extending Splunk Agent Observability with tools designed to connect AI-agent expenditure with business outcomes. Its Tokenomics capability will use usage data from Amazon Bedrock and Amazon CloudWatch to associate each agent’s cost with evaluations of the quality and impact of its work. The service is planned for private preview in December, with general availability through AWS Marketplace in early 2027.
Two additional capabilities—Eval-G1 evaluation models and the Evaluation Tuner—are generally available this month. Eval-G1 models can turn agent traces and expert feedback into custom evaluation models, which Cisco says can run evaluations and guardrails at up to 98% lower cost than commercial large language models, with response times 2.5 times faster. The Evaluation Tuner enables teams to fine-tune evaluator models using operational traces and subject-matter expert feedback.
The third part of the announcement strengthens Cisco IQ’s integration with the Splunk Platform. Cisco Data Fabric will bring vulnerability, lifecycle and configuration intelligence from Cisco IQ into Splunk, enabling security teams to correlate live threat detections with exposed network infrastructure. Cisco IQ has close to 17,000 customers across more than 22 million devices.
New Model Context Protocol and ServiceNow connectors will make infrastructure context available to supported AI agents and IT service management platforms. Exposure findings can be converted into asset-linked ServiceNow problem records, helping coordinate remediation. Cisco is also introducing an AI-powered resilient infrastructure playbook covering critical lifecycle vulnerabilities, known defects and quantum-safe cryptographic readiness.
“Bringing Cisco IQ vulnerability insights into Cisco Data Fabric powered by the Splunk Platform gives analysts that context in the middle of an investigation, so they can decide what to do next,” said Liz Centoni, EVP and Chief Customer Experience Officer, Cisco.
The announcements reflect a broader shift in enterprise security: as AI agents gain the ability to act across systems, organisations need more than model-level safeguards. They must test agent behaviour, monitor quality and expenditure, understand infrastructure exposure and connect security findings to remediation workflows.
Cisco’s approach is to link these functions into a continuous operating loop—testing AI systems before deployment, observing their activity and value in production, and using live infrastructure intelligence to reduce risk.