Cisco adds Anthropic's Claude Mythos to its AI red-teaming tool

Cisco announced Monday that it’s adding Anthropic’s Claude Mythos model to AI Defense, the red-teaming framework it uses to let enterprise security teams stress-test their AI applications and agents, giving those stress tests the benefit of Mythos’s frontier-level reasoning about how to plan an attack. Cisco launched AI Defense in January 2025, and the company says hundreds of customers have already used it to catch dangerous gaps in their AI models, apps and agents before deployment; adding Claude Mythos is meant to let the tool plan more realistic, multi-step attacks and turn the results into concrete attack ideas for each system being tested.

Cisco built in several constraints on how Claude Mythos can be used inside AI Defense: it can only run against targets a customer has explicitly registered and authorized, every red-teaming session is cryptographically tied to a named, authenticated operator, engagements are time-bound and can be halted on demand, and customers interact with the model only through AI Defense’s own structured interface rather than open-ended prompting. “Enterprises are increasingly building AI agents into their core operations and want to test those agents thoroughly before putting them into production,” said Michael Moore, Anthropic’s cybersecurity lead. “By bringing Claude Mythos to Cisco AI Defense, Cisco can help security teams red team their agents with advanced reasoning that plans tests and generates attack ideas for each system.”

Cisco is pairing the red-teaming update with new cost and quality tools for Splunk Agent Observability, including a feature called Tokenomics that ties AI spending data from Amazon Bedrock and CloudWatch to quality scores for an agent’s output, due in private preview in December and generally available on AWS Marketplace in early 2027. The company points to a Splunk research report from last month finding that 81% of IT operations and engineering professionals report tighter cooperation between their security and observability teams as AI agents spread through enterprises.

Why it matters

Cisco is betting that enterprises adopting AI agents will pay for a frontier model to attack their own systems before an outside attacker does — and tying that security pitch to Anthropic’s brand, rather than building the adversarial-reasoning model itself, is a bet that the hardest part of AI security is now judgment, not infrastructure.