News

How to secure AI agents at scale: Cisco AI Defense and Splunk

News | 16.11.2026

AI agents now connect to sensitive data and business tools, yet most enterprises cannot test, observe or measure them at the speed they are deployed. Cisco has introduced an always-on trust loop that combines AI red teaming with Claude Mythos in Cisco AI Defense, cost and quality control in Splunk Agent Observability, and live infrastructure exposure data from Cisco IQ. Softprom is the official distributor of Cisco.

What was announced

At Cisco Live APJC in Melbourne on November 10, 2026, Cisco announced that Anthropic's Claude Mythos is being added to the adaptive red teaming framework of Cisco AI Defense. Since its launch in January 2025, AI Defense has helped hundreds of customers find critical flaws in AI models, applications and agents before production.

Splunk Agent Observability gains Tokenomics, which links agent cost to ROI, and Eval-G1 evaluation models that run evals and guardrails at up to 98% lower cost and 2.5x faster than a commercial LLM. Cisco IQ, used by close to 17,000 customers across more than 22 million devices, now streams vulnerability and lifecycle data into the Splunk Platform.

You cannot scale what you cannot secure, govern, or observe.

Jeetu Patel, Cisco President and Chief Product Officer

Why this matters

AI is compressing the window between vulnerability disclosure and exploitation from weeks to hours, so point-in-time audits leave organizations exposed. According to a Splunk research report, 81% of ITOps and engineering professionals see increased cooperation between security and observability teams. For CIOs and CISOs, the announcement means one control loop for testing agents, tracking token spend and linking threat detections to real network exposure. Procurement leaders gain a clearer basis to justify AI budgets with measurable ROI.

Technical details

  • Realistic adversarial ideation: Claude Mythos generates multi-step, contextual attack paths across complex architectures.
  • Contained access: customers use only the structured AI Defense interface, with no open-ended prompt access.
  • Safeguards: registered and authorized targets only, cryptographic attribution to a named operator, time-bound revocable sessions and dual-layer oversight.
  • Audit trails: every testing action is recorded, and results feed follow-up attack loops.
  • Tokenomics: pairs agent cost from Amazon Bedrock and Amazon CloudWatch with quality evaluations; private preview in December, general availability in AWS Marketplace early 2027.
  • Eval-G1 and Evaluation Tuner: custom small language models built from agentic traces and expert feedback; generally available this month.
  • Cisco Data Fabric integration: Cisco IQ data enters Splunk Enterprise Security via open APIs.
  • MCP and ServiceNow connectors: exposure findings become asset-linked ServiceNow problem records automatically.
  • Resilient Infrastructure Playbook: alerts on known defects and quantum-safe cryptographic readiness.

FAQ

How can we test AI agents before production?

Use adaptive AI red teaming that simulates multi-step attacks against registered targets. Cisco AI Defense with Claude Mythos plans tests and generates attack ideas for each system.

Can Cisco show the ROI of our AI agents?

Yes. Tokenomics in Splunk Agent Observability ties each agent's cost to graded output quality, giving a single view of AI ROI.

When will these features be available?

Eval-G1 models and the Evaluation Tuner are available now. Tokenomics enters private preview in December, and Claude Mythos red teaming is coming soon for select enterprises.

Softprom and Cisco

Softprom is the official distributor of Cisco. Our team helps partners and customers plan, license and deploy Cisco security and observability solutions.

Prepared by the Softprom editorial team based on vendor materials (Softprom DistriFlow). Source: original article.