Cisco plans Claude Mythos integration for AI security testing
Initial access will be limited to select enterprise customers, with tests restricted to systems they own and have registered.
Cisco plans to integrate Anthropic’s Claude Mythos into AI Defense to help businesses test their AI applications and agents for security weaknesses. Announced at Cisco Live in Melbourne, the integration will initially reach select enterprise customers, though Cisco has not set a date.
Mythos will devise attack ideas and testing plans for red teaming, which probes systems for vulnerabilities. Cisco researchers used the model in July to get 15 chatbots to bypass their guardrails.
Customers will not have direct prompting access to Mythos. Instead, the model will perform a defined task on customer-owned systems registered for testing. Cisco said each test will be linked to an identified person and logged for auditing, with the ability to stop it at any point. The model cannot search the broader internet, the company said.
Other security providers are also adopting Mythos. Palo Alto Networks introduced continuous testing using Mythos and OpenAI’s cyber model in September. IBM plans to incorporate Mythos into Concert Protect in early 2027.
Cisco separately announced changes to its Splunk tools for monitoring AI agents’ costs and evaluating their work. Its Tokenomics cost tool is scheduled to enter private preview in December.
