Cisco brings Claude Mythos into its red teaming: the AI attacker becomes a defense tool, on a leash

By TradingView (PR Newswire) · November 9, 2026.

Cisco announced at Cisco Live APJC (Melbourne) that it will integrate Claude Mythos, Anthropic's model, into AI Defense, its AI security platform. The use case: red teaming, that is, simulating attacks against AI applications and agents to find flaws before they reach production. According to Cisco, access will come "soon" and only for select enterprises.

The company says AI Defense, launched in January 2025, has already helped hundreds of customers detect critical flaws. Mythos would contribute adversarial reasoning to plan tests and turn results into attack ideas tailored to each system, with multi-step attack paths.

The most interesting part is the safeguards Cisco describes. The model operates only against targets registered and authorized by the customer, and cannot scan the internet. Each session requires authentication and is cryptographically bound to a named operator. Tests have a fixed objective defined by a person, run for a limited time and can be stopped on demand. Customers have no open access to prompts, only to a structured interface, and every action is logged for auditing. All of this is Cisco's claim; there is no independent verification in the material.

The announcement includes more pieces. Splunk Agent Observability adds "Tokenomics," which cross-references each agent's cost with evaluations of the quality of its work to calculate AI return. It will arrive in private preview in December and reach general availability on AWS Marketplace in early 2027. Cisco is also launching Eval-G1 evaluation models, small language models that, according to Cisco, run evaluations and guardrails up to 98% cheaper than a commercial LLM and 2.5 times faster. Available this month. Finally, Cisco IQ, with about 17,000 customers and more than 22 million devices, will feed vulnerability data into Splunk via MCP (Model Context Protocol, the standard for connecting agents to tools) and ServiceNow connectors.

**Our take.** This is the defensive side of the dual-use problem. The release itself acknowledges that a frontier model's ability to design attacks "cuts both ways" and that AI lowers the barrier to entry for attackers. Cisco's answer is not to limit capability, but to govern who uses it, against what and for how long. It is the approach we advocate: risk is managed through delegated authority, traceability and revocation, not through panic or bans.

Cisco's emphasis on identified operators and bounded sessions points in that direction. And the Tokenomics piece confirms another thesis: the bottleneck for enterprise agents is trust and accountability, not technical capability. Measuring cost against quality is the first step toward scaling sensibly.

The caveats: this is a press release. There are no figures on the effectiveness of red teaming with Mythos, no customer results and no firm availability date beyond "soon." The 98% savings and 2.5x speed figures come from Cisco itself. Independent testing remains to be seen.

In the long term, the trend is encouraging: if the same models that make attacks cheaper can audit systems at machine speed, security stops being a one-off audit and becomes a continuous process. In the short term, the window between a vulnerability's disclosure and its exploitation is shrinking, according to Cisco, from weeks to hours, and companies without this layer will feel it first.