Cisco is bringing Anthropic's Claude Mythos into its AI Defense platform, giving security teams a more powerful way to test whether enterprise AI agents can be manipulated into leaking data, misusing tools, or performing unauthorized actions.
Announced at Cisco Live APJC on Nov. 9, the integration uses Claude Mythos to develop multi-step attacks tailored to individual AI systems. Access will initially be restricted to selected enterprises, and Cisco has not announced a general release date.
The approach targets a growing problem for businesses deploying autonomous agents: security weaknesses that emerge when AI systems interact with internal applications, APIs, and sensitive information. Cisco says its existing adaptive red teaming framework can test these environments, while Mythos will strengthen the reasoning behind its simulated attacks.
How Cisco's Claude Mythos integration works
According to Cisco's announcement, Claude Mythos will augment AI Defense's existing adaptive red teaming system rather than replace it.
Standard red teaming checks common vulnerabilities, including prompt injection, jailbreaks, and sensitive data exposure. Adaptive testing examines a specific application, generates attacks based on its behavior, and changes tactics in response to the results.
Claude Mythos adds more advanced adversarial reasoning to develop contextual attacks involving multiple steps or connected tools.
Cisco is also restricting how customers can use the model. Its technical explanation identifies four safeguards:
- Authorized targets: Tests run only against systems registered and approved by the customer.
- Named operators: Every engagement is linked to an authenticated enterprise user.
- Controlled execution: Testing follows predefined objectives, runs within time limits, and can be stopped.
- Auditability: Actions are recorded for review, while customers cannot freely prompt Mythos or scan arbitrary internet systems.
These restrictions are particularly relevant because Anthropic's Project Glasswing has already demonstrated the model's ability to uncover difficult software vulnerabilities.
The challenge is using those capabilities defensively without giving users unrestricted access to an advanced cybersecurity model.
What eWeek found: Stronger testing, but important gaps remain
Cisco's announcement provides a clearer picture of how Mythos will be controlled than how much it will improve security outcomes.
The distinction between existing capabilities and the upcoming integration is important for enterprises evaluating AI Defense.
| Capability | Availability | What enterprises should know |
|---|---|---|
| Standard AI red teaming | ✔ Available | Tests common AI vulnerabilities |
| Adaptive AI red teaming | ✔ Available | Develops attacks tailored to specific applications |
| Claude Mythos integration | ✖ Coming soon | Adds advanced attack reasoning; selected enterprises first |
| Independent Mythos performance results | ✖ Not published | Detection improvements remain unverified |
| General availability date | ✖ Not announced | Broader access depends on qualification and rollout decisions |
Source: Cisco's Nov. 9 announcement and technical blog.
Three considerations stand out for security teams:
- Testing is not prevention. Discovering an exploitable weakness does not automatically fix it. Organizations still need runtime safeguards, restricted agent permissions, and remediation processes.
- Model capability is not measured by detection accuracy. Cisco has not published comparative results establishing how many additional vulnerabilities Mythos identifies, its false-positive rate, or the time required to validate findings.
- Restricted access limits immediate adoption. Existing AI Defense customers can use standard and adaptive testing, but the Mythos enhancement is not yet broadly available.
The distinction between finding and fixing vulnerabilities is already visible in Anthropic's Mythos research, where large numbers of AI-generated findings created verification and patching challenges.
Cisco also announced related Splunk Agent Observability improvements. Tokenomics, which connects AI spending with output-quality evaluations, enters private preview in December, with AWS Marketplace general availability planned for early 2027.
For enterprises preparing to deploy agents with access to sensitive systems, the practical priority remains testing realistic attack paths, limiting permissions, and verifying that safeguards work before production. Claude Mythos could expand the range of weaknesses those tests uncover, but customer results will be needed to establish the improvement.
Also read: Meta's $300,000 AI agent bug bounty case examines authentication failures, prompt injection, and credential exposure in another enterprise AI deployment.