Cisco Is Using Anthropic's Own Model to Attack Your AI Agents Before Someone Else Does
AI & ML

Cisco Is Using Anthropic's Own Model to Attack Your AI Agents Before Someone Else Does

Claude Mythos is joining Cisco's AI Defense platform as a red-teaming engine that simulates multi-step attacks on enterprise agents, while Splunk ships evaluator models Cisco says cost 98 percent less than a commercial LLM. Together they sketch what AI security tooling looks like once agents are the attack surface.

PublishedNovember 19, 2026
Read time5 min read
Share

What Cisco actually announced

Cisco unveiled a cluster of AI security capabilities on November 9, spanning its AI Defense platform, Splunk's observability stack, and Cisco IQ's device telemetry, all framed around a single argument from President and Chief Product Officer Jeetu Patel: you cannot scale what you cannot secure, govern, or observe. The centerpiece is a partnership with Anthropic that brings Claude Mythos into AI Defense's adaptive red-teaming framework, where it will simulate multi-step attacks against AI applications and agents that customers register, own, and authorize for testing.

Anthropic's Cybersecurity Lead Michael Moore framed the collaboration around a specific enterprise pain point: organizations want to test their AI agents thoroughly before putting them into production, and a frontier model doing the attacking surfaces failure modes that scripted, rule-based red-teaming tools tend to miss. AI Defense itself launched in January 2025 and Cisco says it has already helped hundreds of customers find critical flaws before those systems reached production, giving the Mythos integration an existing customer base to extend into agent-specific testing.

The guardrails Cisco built around an AI that attacks AI

Letting a frontier model run adversarial tests against production systems is the kind of capability that invites an obvious question: what stops the red-teaming agent itself from becoming the incident. Cisco's answer is a structured interface that limits every engagement to targets the customer has explicitly registered, owns, and authorized, removing any ambiguity about scope before testing begins. Each engagement is also tied to a single named human operator, runs against a fixed objective within a time-bound session, and can be halted at any point, with dual-layer oversight operating at both the product and program level.

Every test action gets logged for audit, which matters directly to the compliance teams who will eventually need to explain to a regulator or an internal risk committee exactly what an automated red-teaming exercise did and why. This is the same identity and audit discipline enterprises have been demanding of their production agents applied instead to the tool meant to probe those agents for weaknesses, and it reflects a security industry increasingly unwilling to treat any AI system, offensive or defensive, as exempt from the controls it recommends for everything else.

Why Splunk's small evaluator models matter more than the headline partnership

Buried under the Anthropic news is a detail that will matter more to most IT budgets: Splunk's Eval-G1 models, custom small language models purpose-built for evaluations and guardrails, which Cisco says cost up to 98 percent less than a commercial large language model while responding 2.5 times faster. Eval-G1 and a companion tool called Evaluation Tuner, which turns agent traces and expert feedback into fine-tuned evaluator models, both reach general availability this November, meaning enterprises already running Splunk can adopt them immediately rather than waiting on a future roadmap commitment.

The logic behind small, specialized evaluator models is straightforward once stated: checking whether an agent's output is safe, accurate, or policy-compliant does not require the same scale of model that generated the output in the first place, and running a frontier model as your guardrail on every single agent action is an expensive way to catch a small fraction of genuinely bad outputs. Splunk Tokenomics extends this cost discipline further, tying agent AI spend directly to output quality using Amazon Bedrock and CloudWatch usage data, and enters private preview in December before a planned AWS Marketplace general availability in early 2027.

The device-telemetry layer most of this coverage misses

Cisco IQ, the company's device and infrastructure intelligence layer, now feeds vulnerability, lifecycle, and configuration data from nearly 17,000 customers and more than 22 million devices directly into Splunk Enterprise Security through open APIs. New MCP and ServiceNow connectors convert exposure findings into asset-linked ServiceNow problem records automatically, removing a manual triage step that previously required a security analyst to translate a vulnerability finding into an actionable ticket by hand. Cisco cites a Splunk research figure that 81 percent of ITOps and engineering professionals report increased cooperation between security and observability teams, framing this integration as formalizing a collaboration pattern already underway informally.

An AI-powered Resilient Infrastructure Playbook rounds out the Cisco IQ update, specifically flagging lifecycle vulnerabilities and quantum-safe cryptographic readiness gaps, a forward-looking addition given how far most enterprise infrastructure remains from post-quantum migration. Chief Customer Experience Officer Liz Centoni's framing captures the throughline across all of Cisco's November announcements: knowing where you are exposed only matters if you can actually act on it, which is precisely the gap between vulnerability discovery and remediation that the ServiceNow connector and the Resilient Infrastructure Playbook are both built to close.

What this means for your AI security budget

If your organization already runs Splunk for observability, Eval-G1's general availability this November is worth evaluating immediately against whatever evaluation or guardrail tooling you currently run, given Cisco's cost and latency claims against commercial LLM-based evaluators. Request a side-by-side cost comparison against your current evaluation spend specifically, since a 98 percent cost reduction claim from the vendor itself deserves independent verification against your own agent traffic volume before you treat it as a budget line you can actually bank on.

For the Claude Mythos red-teaming integration specifically, ask Cisco for early access or a detailed briefing on the operator-named, time-bound session model before committing any production AI agent to testing, since the safeguards described publicly are necessarily high level. The broader signal worth internalizing is that AI security tooling is now explicitly agent-aware rather than adapted from traditional application security, and vendors who have not yet articulated an equivalent agent-testing and agent-governed-evaluation story are behind where Cisco, Microsoft, and their peers have already moved.

Tagged#news#ai-ml#ai#llm#agents#agentic-ai#openai#anthropic#regulation#cisco-ai-defense#claude-mythos#splunk-eval-g1#ai-red-teaming#cisco-iq#agent-observability#tokenomics