An AI agent that answers safely in a demo may behave differently when it can reach customer records or trigger a business workflow.
Cisco announced on Nov. 9 that it is adding Anthropic’s Claude Mythos to AI Defense’s adaptive red teaming framework to help test AI applications and agents. The capability is coming soon for select enterprises, giving security partners a potential new tool for assessing customer deployments before they enter production.
For managed security service providers (MSSPs) and systems integrators, the opportunity lies in connecting those tests to a service customers can use: identifying unacceptable agent behavior, investigating findings, and verifying fixes.
What Mythos adds to the testing process
Red teaming deliberately challenges a system to uncover weaknesses. Cisco’s existing standard tests cover common AI risks, including prompt injection, sensitive data exposure, and harmful outputs.
Its adaptive approach lets users define testing criteria in plain language. The framework then develops tailored attacks and adjusts its approach using the target’s responses, according to Cisco’s technical explanation.
Cisco says Mythos will augment the reasoning used to understand a target, plan tests, and generate attack ideas. The model will operate within the testing framework, with a defined job and human-set objectives.
For a partner, a useful assessment should therefore start with a specific business rule. A customer-service agent, for example, might need testing against attempts to retrieve another customer’s records. This is an illustrative scenario.
The distinction matters as providers address gaps in AI agent identity and access governance. Testing needs to reflect the permissions, connected tools, and data boundaries of the actual deployment.
Access comes with strict boundaries
This integration will not give customers an unrestricted Mythos interface.
In its launch announcement, Cisco describes several safeguards:
- Authorized targets: Testing is restricted to enterprise-owned targets that the customer has registered and authorized. It cannot scan arbitrary internet infrastructure.
- Accountable operators: Engagements require authentication and are tied to a named enterprise operator.
- Controlled execution: Sessions have defined objectives, run for a limited time, and can be stopped on demand.
- Recorded activity: Testing actions are logged for audit, with additional safeguards and monitoring around the program.
Partners planning customer engagements should clarify who authorizes each target, who operates the test, and who can stop it. Those responsibilities belong in the scope of work.
Where partners could build a service
Cisco’s announcement supports a potential assessment-and-remediation offering, although providers will need to confirm access and commercial terms before selling Mythos-specific engagements.
A practical service could include three stages:
| Service stage | Partner deliverable |
|---|---|
| Scope the assessment | Inventory the agent’s tools, permissions, and data access; agree on prohibited actions and authorized targets. |
| Validate findings | Review reported weaknesses, document reproducible evidence, and prioritize business impact with the customer. |
| Verify remediation | Retest after changes to permissions, application behavior, or guardrails, and record remaining risk. |
That approach complements Cisco’s broader expansion of partner tools and Cisco IQ access. Providers should still distinguish AI application testing from infrastructure exposure management when explaining what each service covers.
The wider market is moving in a similar direction. CrowdStrike’s latest AI security partnerships extend protection across AI applications, agents, and cloud infrastructure, giving partners more options to evaluate against customer requirements.
Confirm access, then prove the outcome
Cisco says standard and adaptive red teaming are already available in AI Defense. Mythos-powered testing is coming soon for select enterprises, while broader access for qualified customers remains conditional. Dates, scope, and regional availability are subject to change.
Before committing to a customer delivery date, partners should confirm eligibility, regional access, licensing, and whether their intended delivery model is supported. They should also ask where test data is processed, what evidence can be exported, and how customer environments are separated.
Once access is confirmed, start with one agent and one consequential workflow. Agree on the behavior that must be prevented, establish an escalation owner, and define what evidence would demonstrate a successful fix.
A useful customer report should explain what was tested, what failed, what changed, and what remains untested. That gives the customer a basis for a deployment decision and gives the partner a clear standard for follow-up work.
Read more: As partners expand AI security services, rising alert fatigue among MSPs shows why new tools need clear findings, manageable workflows, and human oversight.