A security testing lab.
By Ellis Ward | November 10, 2026
Cisco has a new answer to the question enterprises keep asking about their AI agents: who is testing them before something goes wrong in production? At Cisco Live APJC in Melbourne on Monday, the company said it would bring Anthropic's Claude Mythos into its AI Defense platform to run adversarial red-team tests against customer AI applications and agents.
A Frontier Model on the Attack Side
The core idea is direct: use a frontier model's adversarial reasoning to plan and execute tests against the same AI agents a company wants to protect. Claude Mythos designs the attack scenarios. AI Defense runs them. The customer's agents receive the probing.
Cisco says the arrangement carries strict guardrails. Red-teaming access is restricted to customer-owned, explicitly authorized targets tied to an authenticated operator. Structured access controls and audit trails govern every session. The intent, as Cisco describes it, is to give security teams the same quality of adversarial intelligence that a well-resourced attacker might bring, without opening that exposure to unauthorized parties.
The timing makes sense in context. Cisco launched AI Defense in January 2025 and says it has since helped hundreds of customers find critical flaws in AI applications before those applications reached production. That track record made the internal case for a more sophisticated testing layer.
Red teaming with Claude Mythos is listed as "coming soon for select enterprises." No general availability date has been provided.
Watching Every Token the Agent Spends
The observability side of Monday's announcement may have more immediate reach. Cisco upgraded Splunk Agent Observability to give operations and engineering teams a tighter view of what their agents are doing at runtime.
The company also introduced a tokenomics model that ties each agent's operational cost to evaluations of its output quality. The data sources for that linkage are Amazon Bedrock and CloudWatch. An enterprise running agents on AWS would see per-agent cost figures set against quality scores, surfaced through the same observability stack. Cisco IQ is listed as an integration point, though the announcement does not specify the exact nature of that integration.
Cisco anchors part of the observability pitch to Splunk research. The company cites a survey finding that 81 percent of IT operations and engineering professionals report increased cooperation between security and observability teams. That convergence is part of why the token-cost model sits inside an observability product rather than a standalone security tool. When the people watching agents and the people securing them are increasingly the same people, it helps to put both types of data in one place.
The tokenomics features will enter private preview in December. General availability on AWS Marketplace is scheduled for early 2027.
Eval-G1 Cuts the Per-Evaluation Bill
Alongside the red-teaming and observability updates, Cisco announced Eval-G1, a set of evaluation models designed to run quality checks and guardrails at a fraction of the cost of a commercial large language model.
The mechanics: Eval-G1 takes agent traces and expert feedback as inputs and distills them into custom small models tailored to a specific deployment. Those smaller models then handle the evaluations and guardrails, rather than routing every check through a large commercial LLM. Cisco puts the cost reduction at up to 98 percent compared with the commercial LLM baseline.
For enterprises running agents at scale, the arithmetic matters. If every agent output passes through a quality evaluation step, and that step runs at commercial LLM rates, the bill grows with usage. Eval-G1 is meant to make comprehensive evaluation economically viable rather than a budget line that disappears when volume rises.
Eval-G1 is generally available this month.
What Is and Is Not Ready
The announcement covers three distinct availability windows. Eval-G1 is available now. The tokenomics model enters private preview in December and reaches AWS Marketplace in early 2027. Red teaming with Claude Mythos is coming soon for select enterprises, with no broader release date provided.
That staggered schedule reflects different risk profiles. An evaluation model running on internal data at reduced cost is a relatively contained deployment. A frontier model planning adversarial attacks against enterprise AI, even under strict controls, warrants a more careful rollout to vetted customers first.
The broader picture at Cisco Live APJC is a company treating agent security as infrastructure. The combination of frontier-model red teaming, runtime observability tied to cost data, and cheaper evaluation models represents a layered approach. Each layer targets a different point in the agent lifecycle: pre-production testing, runtime monitoring, and ongoing output evaluation. Whether that layering holds under real-world adversarial conditions will become clearer as the red-teaming service moves beyond its initial select-enterprise stage.
Sources
- Cisco Newsroom: Cisco Unveils Frontier AI Red-Teaming and Agent Observability to Secure the Agentic Enterprise
- PR Newswire: Cisco Unveils Frontier AI Red-Teaming and Agent Observability to Secure the Agentic Enterprise
- Enterprise IT World: Cisco Adds Claude Mythos to AI Defense, Expands Splunk Agent Observability for Enterprise AI Security
