Cisco adds Frontier AI Red Teaming and Agent Observability to Enterprise Security portfolio

Cisco is bringing advanced adversarial testing, real-time agent evaluation, and infrastructure exposure data into its security and operations platforms.
Cisco has announced an expansion of its enterprise AI security portfolio, introducing new red teaming capabilities, agent observability tools, and infrastructure risk integrations at Cisco Live APJC on November 9, 2026.
The announcement brings together Cisco AI Defense, Splunk Agent Observability, and Cisco IQ as enterprises deploy increasingly autonomous software agents across applications and operational systems.
The most notable development is Cisco’s plan to incorporate Anthropic’s Claude Mythos into AI Defense for adversarial testing of AI applications and agents.
Testing agents before production
Cisco’s adaptive red teaming framework is designed to simulate multi-step attacks against enterprise AI systems, including attempts to exploit tool access, application behavior, and interconnected workflows.
The planned Claude Mythos integration would extend this testing through advanced reasoning that generates attack scenarios tailored to individual systems.
Cisco says access will be restricted to customer-authorized targets, with authenticated operators, time-bound engagements, revocable execution, and recorded audit trails.
The integration is coming soon for selected enterprises. Cisco has not announced general availability.
Observability moves beyond token tracking with Cisco AI Defense
Cisco also introduced enhancements to Splunk Agent Observability that connect agent operating costs with output quality.
Its Tokenomics capability is intended to correlate token expenditure with evaluation results using Amazon Bedrock and CloudWatch data. A private preview is scheduled for December, followed by planned AWS Marketplace availability in early 2027.
Meanwhile, Splunk’s Eval-G1 evaluation models and Evaluation Tuner are generally available beginning in November.
Cisco reports that its evaluation models can deliver responses 2.5 times faster and operate at up to 98% lower evaluation cost than commercial large language models. These are company-reported performance figures.
Connecting vulnerabilities to remediation
Cisco IQ is also being integrated with Splunk through Cisco Data Fabric, providing security teams with infrastructure exposure, configuration, and lifecycle information.
Connectors for Model Context Protocol (MCP) and ServiceNow are intended to help translate security findings into operational remediation workflows.
AIAUTOMATEIN analysis: When permitted actions become security failures
The more consequential risk in enterprise AI security is not always a malicious prompt or an unsafe response. It is what happens when an agent combines legitimate permissions, connected tools, and access to sensitive information in ways its operators never intended.
An agent may retrieve authorized records, invoke approved services, and execute a permitted workflow, yet still produce an unauthorized outcome.
Cisco’s planned Claude Mythos integration addresses this problem from an adversarial testing perspective. Rather than relying exclusively on predefined attack categories, adaptive red teaming can examine how an application behaves within its particular operating environment.
The distinction matters. Identifying a prompt injection vulnerability is useful; establishing whether that vulnerability can trigger an unauthorized tool action or expose sensitive information is considerably more valuable.
Cisco’s proposed safeguards also deserve scrutiny. Advanced attack generation will operate within registered targets, human-defined objectives, and auditable testing sessions. These controls are intended to keep the testing process itself from becoming another operational exposure.
What remains unproven is how effectively the forthcoming integration will discover vulnerabilities that existing testing methods miss, and whether those findings will translate into durable production controls.
For enterprise security teams, the measure of success should be straightforward: reproducible findings, traceable remediation, and fewer unauthorized outcomes in production.



