Cisco brings Claude Mythos into controlled AI red teaming

Detailed summary

Cisco is adding Anthropic’s model to tests of enterprise AI applications and agents, with customers setting authorised targets and objectives through a restricted interface.

Cisco is adding Anthropic’s model to tests of enterprise AI applications and agents, with customers setting authorised targets and objectives through a restricted interface.

Cisco is bringing Anthropic’s Claude Mythos into its AI Defense platform to help organisations discover how their AI applications and agents could be manipulated once connected to business tools and sensitive data.

Announced in conjunction with Cisco Live in Melbourne, the integration will add Mythos’ reasoning capabilities to Cisco’s adaptive red teaming engine. The aim is to develop attacks tailored to a particular deployment, exposing weaknesses that generic tests may miss.

Importantly, the customer interaction stays within AI Defense’s structured testing interface. Customers specify an authorised target and testing objectives; Cisco’s engine conducts the engagement within those boundaries. Customers do not themselves receive an open-ended interface through which they can freely prompt Mythos.

Availability is still ahead; Cisco says the Mythos-enhanced capability is coming soon for selected enterprises, while broader availability for qualified customers is in progress.

Red teaming deliberately challenges a system to find failures before an attacker does. For AI, those failures can include instructions that override intended behaviour, disclosure of sensitive information, or actions outside an agent’s permitted role.

Cisco already offers standard and adaptive red teaming in AI Defense. Standard testing covers common issues such as prompt injection, jailbreaks, sensitive data exposure, and harmful content. Adaptive testing takes account of the particular application and the risks its owner wants to investigate.

However, Mythos will greatly strengthen the reasoninv involved in understanding a target and developing creative attack ideas. Cisco ensures the model will operate with a fixed job, a registered target, and objectives defined by a human.

The Cisco team, in conjunction with customers, describe…

Opening of the article, excerpted from iTWire. To read the rest, open the article on the publisher's site.