From the Cisco edition of November 9, 2026
Cisco adds Anthropic Claude Mythos to AI Defense adaptive red teaming, extends Splunk Tokenomics to AWS agents and feeds Cisco IQ exposure data into Splunk (vendor announcement)
Cisco Systems (with Anthropic; Splunk by Cisco) · AI security, agent observability and network lifecycle/vulnerability management; cross-SLED applicability (state/local IT, higher ed, K-12 districts, public safety and HHS agencies building AI agents) · Global announcement at Cisco Live APJC, Melbourne; U.S. availability and regions not specified
- Publisher
- Cisco Newsroom
- Original publication
- November 9, 2026 (U.S. newsroom date; dateline Melbourne Nov. 10)
- Source retrieved
- 2026-11-10
- Event date
- 2026-11-10
What happened
Cisco says AI Defense's adaptive red teaming will add Anthropic's Claude Mythos to generate multi-step, system-specific attacks against customers' own AI apps and agents, restricted to registered and authorized targets, tied to a named operator, time-bound, revocable and audited, with no open prompt access. Mythos red teaming is 'coming soon for select enterprises'. Splunk Tokenomics will pair agent cost (Amazon Bedrock/CloudWatch usage) with quality evaluations and customer KPIs (private preview December, AWS Marketplace GA early 2027); Eval-G1 evaluator models and an evaluation tuner are GA this month. Cisco IQ vulnerability, lifecycle and configuration data will flow into Splunk via Cisco Data Fabric, with MCP and ServiceNow connectors.
Why it matters
SLED agencies and universities piloting citizen-facing or internal AI agents need pre-production adversarial testing and cost accountability. The announcement offers a vendor-managed path to frontier-model red teaming without the agency holding the model, but availability is limited to select enterprises, and no public-sector customer, FedRAMP/StateRAMP status, or test results are given. Cisco IQ-to-Splunk exposure correlation is relevant to state and campus networks with large Cisco estates.
Evidence and measured results
Vendor press release. Vendor-stated figures: AI Defense helped 'hundreds of customers' since January 2025; Cisco IQ close to 17,000 customers across more than 22 million devices; Eval-G1 up to 98% lower cost than a commercial LLM and 2.5x faster; 81% of ITOps/engineering respondents cite more security-observability cooperation (Splunk survey). No benchmarks, sample, methodology or customer outcomes are published.
Limitations and uncertainty
Announcement-only; timelines explicitly subject to change. 'Hundreds of customers' (press release) and '400 to 500' (CRN interview) are vendor numbers. Accuracy claims for Eval-G1 differ across Cisco/Splunk pages ('high accuracy' vs. 'up to 95%'), and Tokenomics preview timing differs between Splunk pages (November vs. December). No public-sector authorization status stated.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-10; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agencies deploying AI agents lack a repeatable, auditable way to adversarially test them before production and cannot connect agent spend to service outcomes.
- Stakeholders
- Agency CISO / campus security officer
- AI program or innovation lead
- Application owners of the agent being tested
- Budget/finance office
- Privacy officer
- Discovery questions
- Which AI agents or apps are moving to production in the next two quarters, and how are they tested today?
- Do you already license Cisco AI Defense, and do your tenants meet your hosting/authorization requirements?
- Which data classifications (CJIS, FERPA, HIPAA, PII) would a red-team test touch?
- Do your agents run on Amazon Bedrock, or on other clouds/on-prem models?
- Who approves adversarial testing against production-adjacent systems?
- Value hypothesis
- For agencies already on Cisco AI Defense and Splunk, adaptive red teaming plus exposure data could shorten pre-production security review for a small set of agents, if results are triaged and remediated.
- Potential engagement
- A scoped AI-agent security readiness review for one or two agents using currently available AI Defense red teaming, with a plan to evaluate Mythos-backed testing if and when access is offered.
- Unsupported claims
- That Mythos-backed red teaming is available now to public-sector customers
- Any detection-rate or vulnerability-reduction figure
- Guaranteed AI cost savings or ROI from Tokenomics
- FedRAMP, StateRAMP or CJIS compliance for these new features
Pre-sales engineering
Role takeaway
- Fit
- Best fit for SLED organizations already using Cisco AI Defense, Splunk Enterprise Security and Cisco IQ, with AI agents hosted on reachable endpoints and (for Tokenomics ROI) on Amazon Bedrock.
- Architecture and integration
- AI Defense SaaS red teams registered targets; Splunk Agent Observability ingests Bedrock/CloudWatch usage; Cisco Data Fabric brings Cisco IQ exposure data into Splunk ES; MCP and ServiceNow connectors pass exposure context to agents and ITSM.
- Prerequisites
- AI Defense license and registered target endpoints
- Splunk Cloud/Enterprise Security for exposure correlation
- Amazon Bedrock and CloudWatch for the Tokenomics ROI view
- Defined agent success KPIs
- Cisco IQ coverage of network estate
- Constraints
- Mythos red teaming limited to select enterprises; dates and regions subject to change
- Tokenomics AWS ROI in private preview December, GA early 2027
- No public-sector authorization status stated
- Non-AWS model estates not covered by the announced ROI view
- Security
- Restrict red teaming to non-production or approved targets, keep audit logs in agency-controlled storage, scope MCP connector credentials to least privilege, and confirm data residency of test traffic.
- Proof of value
- Run standard and adaptive red teaming on one agent and count confirmed, remediated findings
- Compare red-team findings with an existing manual or third-party test
- Verify audit records attribute every test to a named operator
- Check Cisco IQ exposure records reconcile with Splunk ES detections for a sample of devices
Delivery
Role takeaway
- Work
- Pilot planning for AI-agent red teaming and exposure correlation: target registration, rules of engagement, findings triage, and KPI definition for cost reporting.
- Dependencies
- Agency security approval for adversarial testing
- Application owner availability for remediation
- Cisco access to Mythos-backed features (pending)
- Splunk ES and Cisco Data Fabric configuration
- Ownership
- Agency or campus security office owns testing and findings; application owners own remediation; finance/IT governance owns cost KPIs.
- Skills and adoption
- Security staff need training on writing adaptive test criteria and interpreting reports; agent owners need guardrail configuration skills in Policy Studio.
- Governance checkpoints
- Written authorization and target list before first test
- Privacy review when tests touch regulated data
- Go/no-go review of findings before agent production release
- Quarterly review of agent cost vs. agreed KPIs
- Acceptance criteria
- Proposed: 100% of tests attributed to named operators in audit logs
- Proposed: all critical findings triaged within an agreed window
- Proposed: agreed KPI defined for each agent before ROI reporting begins
- Risks
- Preview timelines slip
- Tests disrupt shared systems
- Findings outpace remediation capacity
- Over-reliance on vendor-reported metrics
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Red teaming runs inside Cisco's AI Defense SaaS against customer-registered endpoints, so targets must be reachable and registered; Tokenomics' AWS ROI view depends on Amazon Bedrock and CloudWatch telemetry, so Azure, Google or on-prem model estates are not covered by this specific release. Cisco IQ data lands in Splunk Enterprise Security through Cisco Data Fabric and open APIs; MCP connectors expose infrastructure context to AI agents.
Governance
Who approves, reviews and stays accountable for outcomes?
Named-operator attribution and audit trails support authorization records for testing, but agencies still need their own rules of engagement, data-handling approvals, and change control before red teaming production-adjacent systems. Cost-to-KPI ROI reporting requires agency-defined success metrics.
Security and privacy
What data, permissions and controls need testing?
Adversarial tests may exercise sensitive data paths (PII, CJIS, FERPA, HIPAA data) in the target system; agencies must confirm what test data and responses leave their boundary for Cisco/Anthropic processing. Exposing MCP access to infrastructure data creates a new privileged integration to secure.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Limited direct accessibility content in the source; workforce implication is the need for staff able to configure, interpret and act on AI security or cost tooling outputs (interpretation).
Procurement
What should contracts, pricing and exit terms secure?
Check whether new AI features arrive under existing Cisco/Splunk contracts, whether new subprocessors (e.g., Anthropic) are introduced, and whether public-sector authorizations cover them before enablement.
Operating model
Which teams own the service once it runs?
Security, observability and finance teams need a shared operating loop (test, observe cost/quality, remediate) with named owners; the source does not describe a public-sector operating model.
Publication history
- 2026-11-09Cisco · Issue 106 resources
Stable resource ID: cisco-ai-defense-mythos-red-teaming-tokenomics-cisco-iq-20261109