Cisco plans to use Claude Mythos for controlled red-team testing of AI applications and agents in Cisco AI Defense. IBM plans to incorporate the model into IBM Concert Protect for code analysis, exploitability assessment and remediation support. Cisco’s announcement was dated November 10, 2026; IBM’s announcement describes its plan without assigning it an event date. (Cisco; IBM)

The key distinction is the target: Cisco’s plan focuses on testing AI systems, while IBM’s focuses on finding and addressing weaknesses in application code.

The two planned uses at a glance

PlatformIntended focus
Cisco AI Defense with Claude MythosRed-team testing of AI applications and agents
IBM Concert Protect with Claude MythosCode analysis, exploitability assessment and remediation support

Cisco plans controlled testing of AI applications and agents

Cisco says Mythos is intended to help plan tailored tests and generate contextual, multi-step attack ideas for specific AI applications and agents. Customers would interact through a structured AI Defense interface, rather than through open-ended prompts to the model.

The company describes safeguards around where and how those tests would run. Targets must be registered, owned and authorized by the customer; Cisco says the system cannot scan arbitrary internet infrastructure. Each engagement would require authentication and be tied cryptographically to a named enterprise operator. People would define the objectives and time limits, could stop a session on demand, and would have an audit trail of testing actions.

In its November 10 announcement, Cisco described the capability as coming soon to select enterprises. The plan is for adversarial testing of authorized AI systems, with defined human oversight and limits on the test session.

IBM plans code analysis, exploitability checks and remediation support

IBM describes a three-stage workflow for Concert Protect: Scan, Prove and Fix. Mythos is intended to reason over customer-selected application code to identify weaknesses that lack existing signatures or CVE records. A CVE is a record identifying a publicly known cybersecurity vulnerability.

The planned assessment would use code, configuration and infrastructure context to evaluate whether a weakness is exploitable. That contextual step is meant to distinguish a finding that appears in code from one that presents a practical risk in its particular environment.

For remediation, IBM describes risk-prioritized fixes, including trusted backported open-source fixes and targeted code changes. IBM says developers would validate, review and approve proposed changes through pull-request workflows, with Red Hat Lightwell named as part of the remediation work. The workflow connects discovery to developer-reviewed changes rather than treating a scan result as an automatically approved repair. (IBM)