Cisco unveiled three tools on Tuesday to test, watch and rein in workplace AI bots, including one to simulate attacks, one to track bot actions and new controls to limit spending.

Key Takeaways

  • Cisco rolled out three product upgrades on Tuesday to secure agentic enterprise workflows and control risk
  • Cisco introduced Claude Mythos inside AI Defense to simulate real-world attacks on AI applications and agents
  • Cisco tied the package to spending controls through enhanced tokenomics, making cost part of security policy

Cisco said the changes were introduced at Cisco Live APJC. Cisco sells networking and security software, and it is trying to turn that installed base into a control layer for AI agents.

The release groups the update into one loop.
AI red teaming stress-tests agents before deployment, and Cisco IQ tries to map security findings to real systems that need fixing.

AI Security Cisco Wants One Trust Loop

The most concrete change is Claude Mythos inside AI Defense. Cisco said the model will help simulate real-world attacks on AI applications and agents, giving defenders a stronger way to probe prompt injection, tool misuse and unsafe actions.

Splunk Agent Observability tackles a different problem.
Enterprises can build an agent that works in a lab, then lose visibility once it starts calling tools, moving data and spawning substeps in production.

Cisco also tied the package to spending controls through enhanced tokenomics.
In practice, that means cost is becoming part of security policy. A firm may decide some agent paths are too expensive or risky to run without tighter approval.

Also Read: Supermicro Hands Companies A Memory Saver For Thousands Of Chats

AI Security Moved From Models To Workflows

A year ago, most AI security products focused on a model's raw answers.
Could it produce disallowed text. Would it leak a secret.
Could it be jailbroken.

That frame is no longer enough. Enterprise agents now call databases, ticketing systems and internal tools.
The risk sits in the workflow, not only the words. A well-behaved model can still do damage if the surrounding permissions are loose.

Cisco is entering a crowded field, but it has one advantage.
It already sits near the network, the logs and the identity stack inside large companies. That makes Cisco a natural place to sell AI red teaming as one more operating control rather than a separate experiment.

How AI Red Teaming Changes With Agents

AI red teaming used to mean trying to trick a chatbot into saying something it should not.
For Cisco, AI red teaming now means attacking the whole agent loop, from the prompt to the tool call to the downstream system.

That is why Cisco paired Claude Mythos with Splunk Agent Observability and Cisco IQ. Red teaming generates the bad cases.
Observability records what the agent actually did. Cisco IQ tries to connect that behavior to a router, server or application that a security team can patch or isolate.

For large companies, this matters because agent failures are often operational before they are reputational.
A bad answer is embarrassing. A bad tool action can open a ticket flood, expose customer data or change a system record.

Before this, teams often bought separate tools for testing, logging and remediation.
Cisco is betting enterprises want one vendor to connect all three. The next test is whether customers accept one control plane for both AI risk and AI spend, because the release makes those two budgets look increasingly inseparable.

Read Next: Meta Teams Up With Rival Google To Put Its 3 Muse AI Models In The Cloud

Originally published on Fathom News.