Cisco adds Anthropic's Claude Mythos to AI Defense to test software agents

Cisco has revealed an agreement with Anthropic to integrate the Claude Mythos model directly into its AI Defense system, announcing the rollout on Tuesday at Cisco Live APJC in Melbourne.

The tool is designed to run automated red teaming exercises against artificial intelligence agents. As workplaces shift from basic chat tools to software agents that execute multi-step workflows across internal databases, security teams face new failure points. Mythos will simulate adversarial attacks by reasoning through multi-stage paths to discover flaws before external actors can locate them.

Anthropic does not offer open, unprompted access to Claude Mythos for general end users. Instead, the model operates inside pre-built corporate software environments under explicit guardrails. Cisco stated that the integration restricts Mythos so that it can only evaluate targets explicitly registered and authorised by the client business. Each security test must be tied to an authenticated operator, maintains an audit log, and cannot search or probe across the broader internet.

Alongside the security integration, Cisco detailed updates to Splunk Agent Observability. The company previewed a cost-tracking tool named Tokenomics, which measures the volume of tokens used against actual task output quality, using telemetry from Amazon Bedrock and Amazon CloudWatch. It also released Eval-G1, a set of specialised evaluation models designed to benchmark agent performance at lower computing expenses than commercial frontier models.

What this means for Australian businesses

Businesses deploying autonomous software tools should review how those systems access core data:

• Audit the permissions granted to third-party agents before giving them access to customer files or financial records. Limit credentials to only what the specific task requires.
• Set up basic guardrails and review workflows manually while usage is small. Software agents that connect to internal tools can be manipulated through prompt injections or unexpected input.
• Track token use against clear business results. As automation scales across internal operations, background computing expenses can quietly accumulate without a measurable lift in productivity.