Cisco provides a continuous trust loop to secure and monitor agents by integrating Claude Mythos into adaptive red teaming tests along with Splunk token monitoring tools and 98% cost-reduced evaluation models.
Melbourne conference announcement and Claude Mythos integration into the adaptive testing engine
Sandboxed controls and prevention of direct access or random scanning
Tokenomics tool in Splunk and Eval-G1 models 98% lower in cost
Cisco IQ updates and enterprise availability timeline
Frequently asked questions
Melbourne conference announcement and Claude Mythos integration into the adaptive testing engine
In a statement dated November 10, 2026, during the Cisco Live Asia-Pacific, Japan, and China conference in Melbourne, Australia, Cisco announced that it will add Anthropic’s “Claude Mythos” model to its “AI Defense” security product, according to the official press release in the Cisco Newsroom and the Cisco AI technical blog.
AI Defense offers two types of scanning: standard red team testing covers common vulnerabilities like prompt injection, jailbreaking, and sensitive data exposure, while adaptive testing allows users to describe custom scanning criteria in natural language to detail attacks on a specific system. Mythos will enhance this adaptive engine by generating attack ideas and planning multi-step tests for each target, making the tool “one of the first AI security products” to employ the model in this manner.
Sandboxed controls and prevention of direct access or random scanning
Customers will not be able to direct text prompts straight to the Mythos model; the model operates exclusively against targets registered by the customer that have legal authorization for testing, and it cannot randomly scan internet infrastructure. Furthermore, each test operation is cryptographically tied to a named operator and runs within a time-bound session that can be stopped on demand and is fully logged for auditing.
DJ Sampath, head of AI software and platforms at Cisco, explained that the company runs Mythos inside a sandboxed environment with total control over what it reaches and how its sub-agents are spawned, while Michael Moore, head of cybersecurity at Anthropic, stated: “Organizations are increasingly building AI agents into their core operations and want to rigorously test these agents before putting them into production.”
Tokenomics tool in Splunk and Eval-G1 models 98% lower in cost
Cisco paired the announcement with new agent observability tools in Splunk and Cisco IQ updates described as a continuous “trust loop.” It unveiled the “Tokenomics” cost governance tool in Splunk, which uses Amazon Bedrock and CloudWatch data to compare each agent’s cost against classified evaluations of its work. Private preview will begin in December, with availability via the Amazon Web Services marketplace in early 2027.
It also launched new “Eval-G1” evaluation models, which run evaluations and guardrails at up to 98% lower cost compared to major commercial language models, and are generally available starting this month alongside the evaluation tuning tool.
Cisco IQ updates and enterprise availability timeline
The Cisco IQ platform now links live threat telemetry to exposed network infrastructure to accelerate remediation, and connects with ServiceNow to turn vulnerabilities into IT service tickets.
Adaptive testing with the Claude Mythos model will be available “soon” for select enterprises without a final deadline, with broader availability planned for qualified customers. Cisco researchers previously used Mythos last July to prompt 15 chatbots to bypass their guardrails, and AI Defense has helped hundreds of customers discover vulnerabilities prior to deployment since its launch in January 2025.
The package of announcements from Cisco Live in Melbourne demonstrates that securing agent-dependent enterprises requires combining proactive offensive testing with continuous cost and performance monitoring. Integrating the Claude Mythos model within isolated and cryptographically authorized controls, alongside Splunk tools and low-cost evaluation models, gives cybersecurity teams a practical framework to discover vulnerabilities and control spending before expanding production. This blend of controlled offensive scanning and token consumption governance enhances the ability of large enterprises to scale agent deployments securely without losing security or financial control.
Frequently asked questions
Question: How will the Claude Mythos model be used inside the Cisco AI Defense platform?
Answer: It will be integrated into the adaptive red teaming test engine to plan and simulate complex, multi-step attacks against enterprise-registered AI applications and agents prior to production.
Question: Can Cisco customers chat with the Claude Mythos model directly or scan external websites?
Answer: No; the model operates in a sandboxed environment exclusively against registered and authorized targets, and is cryptographically tied to a specific operator within a timed and audited session.
Question: What are the Tokenomics tool and Eval-G1 models announced by Cisco?
Answer: Tokenomics is a Splunk tool that compares token consumption cost to agent output quality, arriving in early 2027, while Eval-G1 models—available this month—reduce evaluation costs by up to 98%.
Question: When will the Claude Mythos integration in AI Defense become available to businesses?
Answer: Cisco stated it will be available soon for a select group of qualified enterprises without a specific release date, followed by wider availability according to qualification terms.