Cisco Embeds Anthropic's Claude Mythos in AI Defense for Select Few

Cisco has embedded Anthropic's Claude Mythos model into its AI Defense platform, but only a select group of enterprises will get access for now. The company announced the capability at Cisco Live Melbourne 2026, positioning it as a way for businesses to find security vulnerabilities in AI agents before attackers can exploit them.

The exclusivity is deliberate. Anthropic has kept Mythos in very few hands, fearing that the model's purported bug-finding abilities could fall into the wrong hands. A small number of firms, Cisco among them, tested the model as part of Project Glasswing. Cisco is now carrying that restricted posture into its commercial release, saying broader availability would come to "qualified customers on a when-and-if available basis." In its Australia announcement, the company did not specify which kinds of business qualify for the upgrade.

What does the restricted design look like?

The technical claims around the integration center on control and accountability. Cisco said AI Defense operates exclusively against registered, authorized targets, and does not arbitrarily scan internet infrastructure. That is a meaningful distinction for a product built on a model whose core strength is finding exploitable flaws.

Red teaming engagements will require authentication, according to Cisco, and each session will be cryptographically tied to a named enterprise operator. In practice, that means every simulated attack traceable to an identifiable person inside the customer organization rather than an anonymous automated run.

Cisco also outlined two further safeguards. The first is revocable execution: engagements run with fixed objectives defined by a human-in-the-loop user, in time-bound sessions that an operator can halt on demand. The second is dual-layer oversight, which couples product-level controls with program-level safeguards and monitoring.

The company claimed its offering is one of the first AI products on the market to deliver a Mythos-driven adaptive AI red teaming framework. That claim rests on the model's scarcity rather than any benchmark, since only a handful of firms have had access to Mythos at all.

Why do enterprises want this?

The commercial context is the rapid spread of AI agents into production systems. Michael Moore, cybersecurity lead at Anthropic, framed the demand directly.

"Enterprises are increasingly building AI agents into their core operations and want to test those agents thoroughly before putting them into production," Moore said. "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."

The emphasis on pre-production testing matters. Red teaming an agent before deployment is cheaper than remediating a breach after the fact, the same way a structural inspection costs less than a collapse.

How much is marketing?

Some care is needed in separating substance from vendor positioning. The safeguards Cisco described — authentication, named operators, human-in-the-loop objectives, revocable sessions — are concrete, verifiable design choices that respond directly to Anthropic's stated concern about misuse. They represent the commercial compromise that made the deal possible at all.

By contrast, the "first on the market" framing is a vendor marketing claim about a field with almost no competitors, given Anthropic's own restrictions on who may use Mythos. The announcement also leaves commercial questions open: no pricing, no named customers, no eligibility criteria and no dates beyond the vague promise of broader availability to qualified customers.

What is clear is the direction. Cisco is betting that security testing for AI agents becomes a standard procurement category, and that privileged access to a scarce, high-capability model is a differentiator rivals cannot easily copy. Until Anthropic widens the circle, the Melbourne announcement serves as much as a statement of access as of product — a reminder that in the current market for frontier security models, who gets to deploy a tool matters nearly as much as what the tool can do.