AI agents need a security rehearsal

An AI agent approves a refund outside company policy. Would your team catch the weakness in testing, or would a customer find it first?

Red teaming is a controlled security exercise. An authorized team thinks like an attacker and tests whether an AI agent follows business rules, protects sensitive data, and stays within its permissions.

At Cisco Live APJC in Melbourne on November 10, Cisco announced support for Anthropic’s Claude Mythos in the existing Cisco AI Defense red teaming framework.

Cisco, Microsoft, Google, CrowdStrike, and Palo Alto Networks were among the first companies to use Claude Mythos Preview for defensive security when Anthropic launched Project Glasswing in April 2026.

Cisco’s announcement brings Mythos into its customer-facing red teaming tools, where enterprises will use the model to test their own AI agents.

Cisco says the integration will be available soon to select enterprises.

My takeawayMake red teaming a release requirement for AI agents with permission to take consequential actions.

• Start with the actions with the biggest business impact.
• Assign an owner and a deadline to each finding.
• Retest after fixes or changes in permissions.