Cisco is integrating Anthropic’s Claude Mythos 5 into its AI Defense platform for adaptive red teaming, a form of security testing that develops attacks suited to the system being examined.
Anthropic had limited access to Mythos 5 to a small number of companies because of its potential dangers. The Mythos models are intended to be particularly effective at finding and exploiting security vulnerabilities. Their predecessor, Claude Mythos Preview, was initially made available to selected companies under the Glasswing project to help them proactively close zero-day vulnerabilities. Those partners later received Mythos 5 as an upgrade.
The integration is intended to let customers test AI applications and agents, including their own products, for weaknesses. Users will have controlled access to the model rather than unrestricted access through an interface. Safeguards are intended to prevent misuse.
Cisco says conventional red-teaming methods cannot fully test the risks specific to current AI applications. It has therefore developed and introduced an adaptive agent framework in AI Defense. Unlike standard tests, which check for issues such as prompt injection and disclosure of sensitive data, the adaptive approach is intended to analyse the target environment and devise tailored attack scenarios.
Mythos uses reasoning capabilities to generate attack ideas and test payloads, adjusting its approach in response to the target system. Tests are limited to registered, authorised targets and assignments defined by the user. Activities are attributed to a named user and logged for audits. Cisco says access to the public internet outside the authorised test environment is excluded.
AI Defense’s existing standard and adaptive red-teaming functions are already available. Wider access to the Mythos extension is planned for approved customers, but Cisco has not specified which customer groups will qualify.