AI red teaming is moving up the enterprise agenda as companies hand more work to AI agents that can reach into internal tools and sensitive data. Cisco said on Tuesday (10 November) at Cisco Live APJC in Melbourne that it is adding Anthropic’s Claude Mythos model to Cisco AI Defense, giving security teams a stronger way to stress-test those agents.
The AI red teaming update arrives alongside new Splunk agent observability features and expanded Cisco IQ tools, which Cisco frames as one loop to find risk, measure cost and prove trust. It adds to a wider industry response to AI-era threats, including Sophos’s latest firewall refresh.
Why AI Red Teaming Matters as Agents Multiply
Every tool or database an AI agent connects to adds a possible point of attack. In a global Splunk study released last month, 81% of IT operations and engineering professionals said cooperation between security and observability teams has increased.
Cisco launched AI Defense in January 2025. The company says the platform has since helped hundreds of customers catch critical flaws in AI models, applications and agents before they went live. Its adaptive AI red teaming lets security teams build custom, multi-step attacks against their own systems to surface specific risks.
Cisco says Claude Mythos will plan the tests that matter most for each system and turn the results into tailored attack ideas.
“You cannot scale what you cannot secure, govern, or observe,” said Jeetu Patel, Cisco’s President and Chief Product Officer.
The new AI red teaming capability includes:
- Realistic adversarial ideation: multi-step, context-aware attack paths designed to expose subtle weaknesses in complex architectures.
- Contained access: customers use the model only through AI Defense’s structured interface, with no open-ended prompt access.
- Audit trails and repeatable results: each testing action is logged, and findings can feed into follow-up rounds of testing.
Guardrails Built Into Cisco’s AI Red Teaming
A model that can plan attacks could also be misused, so Cisco has placed several limits around it. Testing runs only against targets the customer has registered, owns and authorised, which means it cannot scan arbitrary internet infrastructure. Each engagement requires authentication and is cryptographically tied to a named operator.
Sessions are time-bound, follow objectives set by a human user and can be halted on demand, with product-level controls running alongside program-level monitoring.
Michael Moore, Cybersecurity Lead at Anthropic, said enterprises want to test agents thoroughly before production, and that the integration lets security teams carry out AI red teaming with reasoning that plans tests and generates attack ideas for each system.
Tracking What AI Agents Cost and Deliver
Security is one side of the trust problem. The other is whether AI spending pays off. Cisco is expanding Splunk Agent Observability with three additions:
- Tokenomics: a cost-governance tool that pairs each agent’s spend, drawn from Amazon Bedrock and Amazon CloudWatch data, with graded evaluations of its output. It extends token-spend tracking launched at .conf26 for agents and coding tools such as Claude Code, Codex and Cursor.
- Eval-G1 evaluation models: small language models built from agent traces and expert feedback. Cisco says they run evaluations and guardrails at up to 98% lower cost than a commercial LLM, with responses 2.5 times faster.
- The Evaluation Tuner: turns agent traces and subject-matter-expert feedback into fine-tuned evaluator models that enforce quality standards.
Cisco IQ Moves From Exposure to Action
Cisco says the gap between a vulnerability being disclosed and exploited has shrunk from weeks to hours. In response, it is feeding real-time data from Cisco IQ, which covers close to 17,000 customers and more than 22 million devices, into enterprise security tools.
A Cisco Data Fabric integration brings that data into the Splunk Platform, so Splunk Enterprise Security analysts can match live threats with network exposure. New Model Context Protocol (MCP) and ServiceNow connectors turn findings into asset-linked problem records, while an AI-powered playbook flags lifecycle vulnerabilities and quantum-safe cryptographic readiness.
Liz Centoni, Cisco’s EVP and Chief Customer Experience Officer, said exposure data only matters if teams can act on it, and that Cisco’s partners and professional services teams aim to help customers move from spotting risk to reducing it.
Availability
Adaptive AI red teaming with Claude Mythos is coming soon for select enterprises. Tokenomics enters private preview in December, with general availability on AWS Marketplace planned for early 2027. The Eval-G1 models and the Evaluation Tuner are generally available from this month. Timelines may change.
For organisations in Malaysia and across the region rolling out AI agents, the direction is clear: AI red teaming, cost tracking and remediation are becoming one workflow rather than separate checks. More details are available on Cisco’s AI Defense page.
