Cisco will add Anthropic's Claude Mythos to its AI Defense product so customers can red-team their own AI apps and agents.
Red teaming means attacking your own system on purpose, before someone else does, to find where it breaks. Here the model plans those attacks. The targets are the customer's own AI apps and agents (software that takes actions on a user's behalf), and the customer is the one running the test.
Cisco attached firm limits, according to the company's announcement. Customers get no open prompt access, so they cannot simply type anything they like into the model. Every test is logged to a named person. The feature goes first to select enterprises rather than to every AI Defense customer.
The rest is our read, not Cisco's stated rationale. A model that can plan attacks cuts both ways, and the restrictions look designed to keep it from becoming a general-purpose attack tool. A named audit trail would also make it harder for a compromised or misused account to run tests unnoticed. Both points are inference, since the source material we have does not explain why Cisco chose these controls.
The approach fits a wider pattern. Security teams have long used scripted tools and human testers for red teaming, and putting a frontier model in that seat is a natural next step as companies deploy agents that act on their own. The notable choice here is less the model than the leash: gated access, named accountability and a staged rollout. Those terms say how Cisco thinks about handing offensive capability to customers, and whether other vendors copy them will say whether it becomes the norm.