Cisco to Add Claude Mythos to AI Defense for Controlled Red Teaming

Cisco announced on November 9, 2026 that it will add Anthropic’s Claude Mythos to AI Defense’s adaptive red-teaming framework for testing AI applications and agents. The announcement at Cisco Live APJC in Melbourne also introduced Splunk agent-observability updates and Cisco IQ integrations, extending the release beyond testing into agent-cost evaluation and infrastructure remediation.

Customers will not be able to prompt Claude Mythos freely. Instead, the model will operate within a fixed job with a registered target and human-defined objectives, supplying reasoning for test planning and attack ideation while AI Defense conducts the engagement.

Red Teaming Stays Inside Authorized Boundaries

A Cisco technical overview distinguishes standard and adaptive testing. Standard red teaming checks common risks such as prompt injection, jailbreaks, sensitive-data exposure and harmful content across thousands of categories, content types and adversarial techniques. Adaptive testing begins with customer criteria written in plain language, analyzes the target and develops tailored attacks, then uses the responses to refine its approach or change tactics before reporting findings and recommending follow-up actions.

The safeguards constrain where that reasoning can be used. Claude Mythos will operate only against targets explicitly registered, owned and authorized for testing by the enterprise customer, and Cisco says it cannot scan arbitrary internet infrastructure. Every engagement requires authentication and is cryptographically tied to a named operator. Sessions are time-bound and can be halted, while each action is recorded for audit and results can inform repeatable follow-up attack loops.

Standard and adaptive red teaming are already available in AI Defense, but the Claude Mythos version is coming soon for select enterprises. Broader availability is planned for qualified customers without a committed date, and its timing, scope and regional availability may change. Cisco reports that AI Defense has helped hundreds of customers identify flaws in models, applications and agents before production since its January 2025 launch.

Splunk Links Agent Cost to Output Quality

The Splunk Agent Observability updates address the operating side of agent deployment. Enhanced Tokenomics will use Amazon Bedrock and Amazon CloudWatch usage data to pair each agent’s cost with evaluations of the quality and impact of its work. Those technical metrics can be combined with customer-defined measures such as Net Promoter Score or cost per claim, extending existing token-spend attribution across AI agents and coding tools including Claude Code, Codex and Cursor.

Eval-G1 uses custom small language models to run evaluations and guardrails. Cisco reports that the models can do so at up to 98% lower cost than a commercial LLM and deliver evaluation responses 2.5 times faster than traditional LLM judges. The evaluation tuner turns agentic traces and subject-matter-expert feedback into fine-tuned evaluators that enforce quality standards and steer or block inaccurate agent behavior.

Availability differs across the observability features. Eval-G1 and the evaluation tuner are generally available starting November 2026. Tokenomics is scheduled for private preview in December 2026, followed by general availability in AWS Marketplace in early 2027.

Cisco IQ Connects Exposure to Remediation

Cisco reports that IQ serves close to 17,000 customers across more than 22 million connected devices, providing asset, coverage and advisory context. The Cisco IQ remediation workflow sends vulnerability, lifecycle and configuration data through Cisco Data Fabric into Splunk Platform via open APIs. Security teams and AI agents can then correlate live detections in Splunk Enterprise Security with underlying network exposure.

Model Context Protocol and ServiceNow connectors add structured infrastructure context to supported agents and IT-service workflows. Exposure findings can be converted into asset-linked ServiceNow problem records without manual data reconciliation. Cisco cautions that finding an exposure does not establish that an attacker exploited it: the workflow is intended to add investigation context and coordinate remediation, not treat each finding as a confirmed breach.

The AI-Powered Resilient Infrastructure Playbook also surfaces critical lifecycle vulnerabilities, personalized alerts for known defects and quantum-safe cryptographic readiness so teams and certified partners can prioritize modernization. Across the package, attack generation stays within authorized jobs, evaluation models can apply guardrails, and exposure data informs investigations without establishing compromise by itself. For now, the evaluation models and tuner are generally available, while Mythos-enabled red teaming and Tokenomics are not, so the end-to-end workflow Cisco describes still depends on future or limited releases.