Cisco adds Anthropic Claude Mythos to AI Defense for red teaming

The evolution of AI Defense

Cisco has spent several years building out its AI security industry presence. It launched AI Defense as the first enterprise offering to combine asset visibility, supply chain scanning, and algorithmic red teaming into a single tool. This unified approach helps organizations manage the risks of deploying artificial intelligence.

As the industry shifted toward agentic technologies, Cisco introduced adaptive red teaming and guardrails. These features allow customers to test for risks that are unique to their specific industries or internal applications. The goal is to move beyond generic security checks to address specific organizational needs.

The company is now extending this framework by incorporating Anthropic’s Claude Mythos. This move is intended to help businesses keep pace as they begin to connect AI agents to complex systems of tools and sensitive resources. Security teams need to understand how these agents behave in interconnected environments before they go live.

Standard vs. adaptive testing

Cisco AI Defense currently offers two distinct types of red team testing. Standard red teaming looks for common vulnerabilities that almost any AI application might face. These include things like prompt injection, jailbreaks, and the generation of harmful content.

Standard tests can cover thousands of different categories and adversarial techniques in just a few minutes. These results are mapped to established safety frameworks like OWASP, MITRE, and NIST. However, these tests are often too broad to catch the specific risks of a unique company deployment.

To solve this, Cisco developed an adaptive red teaming framework. Users can provide custom testing criteria in plain language, and the system analyzes the target to develop tailored attacks. It then refines its tactics based on the target's responses to find deeper vulnerabilities.

How Claude Mythos fits in

The integration of Claude Mythos specifically augments the engine where an attacker’s reasoning is most important. As the system plans tests, the model produces the actual attack ideas. When those ideas become concrete payloads, the capabilities of Claude Mythos provide a sharper edge for the simulation.

Cisco designed this deployment to operate under strict product-level controls. The model sits inside AI Defense with a fixed job and objectives defined by a human user. Customers do not get an open interface to prompt the model freely, which ensures the activity remains contained.

The system is restricted to only the targets a customer has registered and authorized for testing. Every action taken by the model is recorded for audit and attributed to a named operator. This ensures that the high-level reasoning of the model does not compromise the safety of the customer's environment.

Impact on enterprise security

Enterprises across many different industries are beginning to build AI agents into their core operations. Because these uses are so diverse, security teams need a way to find behavioral issues and brand risks early. Claude Mythos helps by providing advanced reasoning that can plan complex tests for these specific systems.

As Michael Moore, Cybersecurity Lead at Anthropic, noted, the goal is to help security teams red team agents with advanced reasoning that plans tests and generates attack ideas. This allows for a more thorough evaluation before an agent is put into production for real-world use.

By making these frontier capabilities available to defenders under verified access, Cisco is shifting how security is handled. Instead of withholding powerful models from everyone, they are providing a controlled way for defenders to use them. This helps organizations identify risks specific to their own deployments before an actual attacker finds them.

Availability and next steps

While standard and adaptive red teaming are already available in Cisco AI Defense, the Claude Mythos integration is coming soon. It will initially be available for select enterprises before a broader rollout to qualified customers. Specific dates and regions for this availability are still subject to change.

Cisco intends to share what they learn as these red teaming engagements progress. The company views this as an important step toward ensuring that the tools used to test AI are as capable as the agents themselves. As agents become more sophisticated, the testing frameworks must evolve alongside them.

The broader goal is to provide a comprehensive experience for finding safety and security vulnerabilities. By combining high-level reasoning with robust guardrails, Cisco aims to give security teams a clear picture of their AI risks. This ongoing development will help organizations adopt agentic technologies with more confidence.