Cisco will embed Anthropic's Claude Mythos into AI Defense to plan attacks against customer AI applications and agents. The integration was announced at the Cisco Live event in Melbourne and will first reach select enterprise customers, with no broad release date set. Mythos will generate attack ideas and test plans, extending a red-teaming practice Cisco already used in July against 15 chatbots. For businesses deploying agents in core operations, this turns adversarial testing into a built-in procurement option.
How Mythos will work inside AI Defense
AI Defense attacks a company's own AI systems to expose weak spots before hackers find them. Within that workflow, Mythos takes the planning role: it designs test scenarios and proposes attack ideas aimed at registered AI apps and agents owned by the customer. Cisco researchers applied the same approach in July, pushing 15 chatbots past their guardrails to study failure modes. The model is not offered as a general chatbot inside the product, but as a bounded testing engine.
The deployment is deliberately constrained. Customers receive no open prompt access to Mythos, and each engagement runs only as a fixed job against systems the customer owns and has registered for testing. Every test is tied to a named person, recorded for audit, and can be stopped at any time. Mythos cannot scan the wider internet, according to a Cisco blog post. The design separates authorized self-testing from open-ended offensive capability.
The move fits a wider vendor race around the same model. Palo Alto Networks launched always-on testing based on Mythos and OpenAI's cyber model in September, while IBM said on Monday it plans to add Mythos to its Concert Protect tool in early 2027. Anthropic distributes Mythos inside partner products rather than handing out the model, and released Mythos 5.2 on Monday for vetted organisations only. Tight controls carry a concrete reason: in April, an unauthorised group reached Mythos through a contractor's access.
What this means for companies running AI agents
For companies putting AI agents into production, red teaming becomes easier to buy than to build. Instead of assembling an internal adversarial team, a Cisco customer can order structured attacks on its own registered systems and receive logged, attributable results. That matters most where agents handle operations, customer data, or transactions, and failures carry direct cost. Small firms gain access to methods previously limited to large security teams, while large firms can standardize testing across many agents.
The limits define the purchase decision. Coverage applies only to owned and registered systems, with no internet-wide scanning and no direct prompting, so unregistered shadow tools and third-party integrations stay outside the test scope. Audit logging to a named operator helps with accountability, but buyers still need to clarify data handling, test depth, remediation reports, and stop procedures. The news alone does not prove that tested agents are safe, only that defined attack scenarios were executed under controls.
Cisco also updated Splunk tools that track what AI agents cost and grade the quality of their work, with a cost tool called Tokenomics entering private preview in December. The signal to watch is availability: when Mythos-based testing moves beyond select enterprises, plus IBM's planned early-2027 integration and adoption of Tokenomics previews. Broader rollout would show whether partner-embedded red teaming becomes a standard part of enterprise AI deployment.
