SI Agent News
DevelopingClaudeAgentsSecurity

Cisco integrates Anthropic Claude Mythos for AI red teaming

Cisco is adding a new security feature to its AI Defense tool that uses Anthropic's Claude Mythos to stress-test corporate AI agents against sophisticated attacks.

The short version

Why it matters

As companies move AI agents into core business processes, they need a way to find flaws before these systems cause real-world damage. This tool provides a structured way to simulate high-level threats without letting the testing AI run wild.

Testing AI against AI

Cisco is expanding its AI Defense security solution to include a new red teaming capability. Red teaming is a security practice where an attacker simulates a real-world cyberattack to find weaknesses in a system. By integrating Claude Mythos from Anthropic, Cisco aims to help businesses test their own AI agents against sophisticated frontier attack techniques.

The primary goal of this integration is to provide a rigorous testing ground for companies that are building AI applications for their core business processes. Many organizations want to ensure their systems are safe before they move them into production environments. This tool allows them to see how their agents might react to complex, multi-step prompts or malicious inputs.

As Michael Moore, head of cybersecurity at Anthropic, noted, companies want to test these agents thoroughly before they go live. The inclusion of Claude Mythos provides security teams with advanced reasoning mechanisms. These mechanisms help the system plan tests and generate specific attack ideas tailored to each unique system being audited.

Identifying behavioral risks

The new adaptive red teaming feature does more than just look for simple security holes or code vulnerabilities. It is designed to identify broader behavioral problems that could arise when an AI agent interacts with the public or internal data. This includes identifying risks to a company's brand and other context-dependent issues that might not be obvious during standard development.

Because AI agents can often be unpredictable, understanding their behavior in various scenarios is critical for safety. The tool helps security professionals see if an agent might provide inappropriate information or behave in a way that contradicts company values. By simulating these scenarios, teams can refine the agent's instructions and guardrails more effectively.

This proactive approach allows for a more comprehensive safety check than traditional manual testing. It moves the needle from simply checking if a system is 'broken' to understanding how it might be manipulated or misused in a real-world context. This is particularly important as AI agents become more autonomous in handling customer-facing tasks.

Strict human oversight

To ensure the testing remains safe and controlled, Cisco has implemented several layers of human oversight. The system does not provide an open interface to Claude Mythos for the general public. Instead, the model is only allowed to address specific goals and targets that are registered, managed, and authorized by the customer for testing purposes.

Every red teaming engagement requires a human user to define the specific mission and the intended targets. This means the AI cannot wander off-task or attempt to attack parts of the infrastructure that were not explicitly approved for the test. This structure ensures that the testing process remains a controlled experiment rather than an unguided exploration.

Security is further bolstered by strict authentication requirements for every session. Each test is cryptographically tied to a specific named individual within the company who is responsible for the actions. This creates a clear audit trail, ensuring that every move the AI makes during a red teaming exercise can be traced back to a human supervisor.

Logging and session limits

Cisco has designed the AI Defense tool to be highly observable and manageable for security teams. Every single action performed by the AI during a red teaming session is logged. This allows administrators to review exactly how the agent attempted to bypass security measures or reach its goals during the test.

The sessions are also designed to be time-limited to prevent any prolonged or unintended activity. Human operators have the ability to stop the session at any time if they notice the AI behaving in an unexpected way or if the test has reached its conclusion. This gives the security team ultimate control over the lifecycle of the attack simulation.

By combining these logging features with time limits and human-defined goals, Cisco provides a sandbox environment for stress-testing. It allows companies to push their AI agents to their limits while maintaining a high degree of safety and accountability. This balance is essential for organizations that are eager to adopt AI but are wary of the security implications.

Availability and rollout

Cisco has announced that the adaptive red teaming feature using Claude Mythos will be available to selected companies in the near future. This initial rollout will likely serve as a pilot for large enterprises that are currently deploying high-stakes AI agents into their production workflows.

While the feature is coming soon, Cisco has not provided a specific date for a wider regional or global release. The company is currently focusing on the initial integration and the specific workflows that allow customers to define and manage their test targets. This phased approach allows for refinement based on early user feedback.

As the technology matures, it is expected that more organizations will seek out these types of automated stress-testing tools. The move toward using one AI model to find the flaws in another represents a significant shift in how cybersecurity teams will approach AI safety. It marks a transition toward more dynamic and sophisticated defense strategies.

What to watch

Try it

Check your company's AI policy to see if there is a formal process for 'red teaming' or stress-testing internal AI tools before they go live.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by Computer Weekly · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news