Cisco announced that it is adding Anthropic’s Claude Mythos to its AI Defense platform for adaptive red teaming of AI applications and agents. The announcement was made at Cisco Live APJC in Melbourne on November 10, with the company describing the integration as a way for enterprise security teams to test agent systems using more advanced adversarial reasoning. Rather than giving customers an unrestricted interface to Mythos, Cisco says the model will operate within its testing framework. The product aims to generate attack ideas and adapt tests to the architecture and behaviour of a registered system.
Mythos is used inside a constrained testing workflow rather than as an open chatbot
The safeguards are a central part of the announcement. Cisco says testing is limited to targets that the customer has registered and is authorised to assess. Each engagement is tied to a named operator, runs with fixed objectives and a time-bounded session, and can be stopped. The company also says activity is recorded for audit. These controls matter because a red-team tool with powerful reasoning could cause harm if it were allowed to scan arbitrary systems or execute actions outside an approved scope. Cisco’s design is an attempt to make advanced testing capabilities available to defenders without turning the product into an unrestricted offensive tool. Cisco’s agentic AI security post describes the operator and target controls behind the testing workflow.
AI Defense already supports standard and adaptive testing. Standard red teaming checks common risks such as prompt injection, jailbreaks, sensitive-data exposure and harmful output. Adaptive testing allows a security team to describe risks specific to its own application, after which the framework develops and refines multi-step tests based on the target’s responses. Adding Mythos aims to strengthen the reasoning used to create those tests. The results still depend on the quality of the target configuration, the criteria selected by the security team and the permissions available to the test agent. A model-assisted test is not a substitute for a complete security review.
Cisco is linking agent security with observability and operating cost
Cisco also announced updates to Splunk Agent Observability and Cisco IQ. The observability changes include evaluator models and a tuning workflow that uses agent traces and feedback from subject-matter experts to build smaller evaluators. Cisco says these can run at substantially lower cost and faster response times than commercial large language models. The company is also connecting threat detections with information about exposed infrastructure and product lifecycle status. The combined message is that enterprises need to know not only whether an agent can be compromised, but also whether it is behaving as expected, using tools appropriately and provideing value relative to its operating cost.
Availability remains limited. Cisco says adaptive red teaming with Claude Mythos will initially be provideed to selected enterprises, with wider access dependent on qualification and later rollout decisions. Other announced observability features have different availability schedules, so they should not be treated as one fully released bundle. Security teams considering the integration should ask how target registration is enforced, how credentials are isolated, what evidence is retained and whether test results can be reproduced. The important development is not simply that a powerful model is being used for security testing; it is the attempt to wrap that capability in an auditable enterprise workflow.
There is also a governance issue in how AI red-team results are handled. A testing platform may identify a prompt-injection path or a tool permission that allows an agent to access information outside its intended role. Security teams need enough detail to reproduce the finding, assign ownership and confirm that the fix works. They should avoid treating a model-generated attack path as proof of a vulnerability until it has been validated against the actual application. Cisco’s emphasis on named operators and recorded sessions addresses accountability, but customers will still need to define retention periods, review findings and ensure that test data does not expose secrets. The quality of the remediation process will matter as much as the model used to discover weaknesses.