Strategic PartnersIssue 0111 records
Cisco
This is the first cisco edition, and it centres on AI for running networks and security operations. At Partner Summit, Cisco launched Model Foundry and reports that a tuned network model found the right fix in 11 of 12 lab problems. It also announced MCP servers and expiring, recorded agent access to network devices, and open-sourced Project Quarry for detecting risky agent behavior. Independent coverage tempers these claims. TechTarget finds network agents remain mostly advisory, ABI says Cisco’s trust controls are asserted rather than proven, and an analyst urges buyers to contract for ownership of tuned models. The SLED deployments in view are in higher education: Internet2 AI PODs and campus SOCs. Splunk AI Assistant reached FedRAMP High.
- Evidence records
- 11
- Cross-source patterns
- 3
- Evidence classes
- 7 vendor claim2 independent reporting1 public-sector association guidance1 independent research
- Outcomes
- 7 emerging3 cautionary1 mixed
- Source freshness
- 8 new this fortnight3 recent
- Research completed
- 2026-10-30
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Cisco’s AI agents for network and security operations are deployed as advisors and triage aids, with humans approving fixes and containment
Across this edition, Cisco’s own showcases and independent reporting describe the same operating model. In the WWT lab test of Model Foundry, the agent recommends a fix for a human to approve. At the .conf26 event SOC, AI agents triaged and correlated alerts while analysts made containment decisions. TechTarget reports that only 18% of Cisco/Omdia survey respondents use agents for remediation and that a Meraki customer is not yet doing agent-based remediation. Cisco’s new Privileged Infrastructure Access, which is still being developed, adds short-lived, recorded agent access to devices. For SLED network and SOC teams, the evidence supports an advisory first phase, not autonomous change.
Operating questionWhich network or SOC tasks will the AI agent only recommend, which will it be allowed to execute, and what recorded approval step sits between the two?
Supporting evidenceCisco launches Model Foundry in Cloud Control for partners; Cisco reports a fine-tuned Deep Network model found the right fix in 11 of 12 WWT lab problemsCisco and Splunk (with Endace and Jamf)Informa TechTarget (reporting); Omdia/Cisco and EMA surveysCisco Partner Summit: MCP servers for networking platforms, governed privileged access for AI agents, ThousandEyes in Wi-Fi 6E/7 access points, and Cisco IQ for partners
Evidence for Cisco’s AI operations and trust claims is vendor-run or vendor-sponsored; independent analysts say proof and ownership terms are missing
The main performance figure this week (11 of 12 lab problems fixed) comes from a five-week Cisco test with one partner, graded against an answer key and by an AI judge. Cisco’s observability findings come from a Cisco-sponsored survey in which 56% of respondents cannot measure LLM response quality. ABI Research says Cisco published no figures on how well its trust controls work and no independent auditor has confirmed them, and an independent analyst calls the lab result a promising pilot that needs testing on the buyer’s own network and contract terms on who owns tuned weights. Public buyers who must show evidence to auditors should ask for their own replay tests and written ownership terms before they rely on these claims.
Operating questionWhat test on our own incidents and data, and what contract language on tuned models and evaluation results, will we require before relying on Cisco AI performance or trust claims?
Supporting evidenceCisco launches Model Foundry in Cloud Control for partners; Cisco reports a fine-tuned Deep Network model found the right fix in 11 of 12 WWT lab problemsShashi Bellamkonda (independent analyst)ABI ResearchCisco / Splunk (survey fielded by Oxford Economics)
Higher education is where Cisco AI infrastructure is most visibly deployed in SLED, but the public evidence is announcements without outcome measures
The two named SLED deployments in this edition are universities: shared Cisco AI PODs at the University of Nebraska-Lincoln and Sacramento State for the National Research Platform, and AI-assisted network operations and a Splunk/XDR SOC at Notre Dame, Auburn and Georgetown. Both are joint or vendor announcements with no utilization, troubleshooting-time or incident metrics, and two of the three campuses in the Cisco feature are private. State, local, public safety, HHS and K-12 deployments did not surface this week.
Operating questionIf we follow these campus examples, what baseline measures (GPU queue time, troubleshooting time, alert handling) will we record before deployment so we can report results afterwards?
Supporting evidenceInternet2 with Cisco; University of Nebraska-Lincoln; California State University, SacramentoCisco (with University of Notre Dame, Auburn University, Georgetown University)
Evidence ledger
Full record · every source keeps its link and limitations
Internet2 and Cisco deploy two Cisco AI PODs at the University of Nebraska-Lincoln and Sacramento State as shared GPU capacity for the National Research Platform
Internet2, through its industry relationship with Cisco, supported deployment of Cisco AI PODs at the University of Nebraska-Lincoln and Sacramento State that join the NSF-supported National Research Platform (NRP), a community-operated pool of compute contributed by over 70 organizations and run as one Kubernetes fabric (Nautilus). Combined, the two PODs provide 8 NVIDIA H200 GPUs with 1.1 TB of GPU memory, 128 CPU cores, 2 TB of system memory and 90 TB of NVMe flash; each site has one Cisco UCS C845A M8 chassis and a BlueField-3 DPU with up to 2x200 Gbps to campus and national R&E networks. Internet2 says the pre-validated design reduces integration risk and setup time and is a repeatable model for other contributors. No utilization, performance or cost figures are reported.
Why it matters, evidence and limitations
- Why it matters
- Shows a concrete public higher-ed path for Cisco AI infrastructure: small, standardized GPU nodes contributed to a shared research platform rather than each campus building its own cluster. Relevant to research universities and regional/state universities with limited budgets.
- Evidence and measured results
- Opened the Internet2 release and read it in full, including the hardware list and the description of NRP and Internet2 network contributions. It is a joint-relationship announcement, so claims about reduced integration risk are the parties’ own and are not measured.
- Limitations and uncertainty
- Joint announcement by a member association and its industry partner; no independent evaluation, utilization data or cost. Published Sept. 22, 2026, before most items in this edition; included as the first cisco-stream baseline for research-computing deployments. Only 8 GPUs in total, so this is a modest increment.
Cisco newsroom feature: Notre Dame, Auburn and Georgetown on AI-assisted network operations, a Splunk/XDR SOC and campus-wide Wi-Fi 7
A Cisco-authored feature describes three university customers. Notre Dame is consolidating nearly 1,000 switches onto a Cisco multi-gigabit platform managed through Catalyst Center, which Cisco says uses AI to monitor network health, trace traffic paths and troubleshoot; Notre Dame is piloting ThousandEyes. Auburn’s cybersecurity operations manager says Splunk, later joined by Cisco XDR, let a relatively small SOC team scale against growing threats and freed time to train student interns. Georgetown’s CIO describes a campus-wide Wi-Fi 7 refresh with over 6,000 access points and 50,000 active devices; Cisco calls it one of the largest Wi-Fi 7 deployments in higher education. No metrics on troubleshooting time, alert volume or incident outcomes are given.
Why it matters, evidence and limitations
- Why it matters
- Named U.S. universities describe AI-assisted network operations and a small-team SOC built on Splunk and XDR, which are common higher-ed staffing problems. Notre Dame and Georgetown are private institutions, so public-university transfer depends on budget and procurement rules.
- Evidence and measured results
- Opened the Cisco Newsroom article and read it in full. Customer statements are quoted by Cisco; no independent measurement or third-party confirmation.
- Limitations and uncertainty
- Vendor marketing feature; quotes selected by Cisco; no quantified outcomes; two of three institutions are private. Treat as deployment signals, not evidence of effectiveness.
Splunk AI Assistant, Edge Processor and Ingest Processor reach FedRAMP High for Splunk Cloud in AWS GovCloud
Splunk says Edge Processor and Ingest Processor have achieved FedRAMP High authorization for Splunk Cloud Platform customers in AWS GovCloud (US), with eligible stacks automatically onboarded to Edge Processor and the Ingest Processor Essentials tier. Splunk AI Assistant, which uses AI to help author and explain SPL queries in natural language, has also achieved FedRAMP High. SPL2 is now available across the platform. Splunk claims filtering, masking and routing data before indexing helps cut costs and protect PII; Edge Processor Essentials is offered at no additional cost. No cost or performance figures are provided.
Why it matters, evidence and limitations
- Why it matters
- The authorization is federal, but state agencies, large counties and public universities running Splunk Cloud in GovCloud can use it as one input to their own security reviews, especially for an AI assistant touching security logs. It does not replace state-specific authorization or CJIS requirements.
- Evidence and measured results
- Opened the Splunk blog and read it in full. The FedRAMP marketplace listing was not separately checked in this run.
- Limitations and uncertainty
- Vendor announcement; authorization status should be verified on the FedRAMP Marketplace; federal scope; no measured outcomes for the AI Assistant.
Splunk State of Observability 2026: 56% cannot measure AI response quality and 55% cannot control AI costs, in a 2,050-respondent survey
Cisco released its sixth Splunk State of Observability report. Oxford Economics surveyed 2,050 ITOps and engineering professionals in 10 countries and 16 industries, including government (federal, state and local) and education, from February through March 2026. Reported results: 87% say monitoring AI is a priority but 54% lack clear business outcomes tied to it; 56% cannot measure AI workload metrics such as LLM response quality; 85% call hallucination visibility important but only 58% say they are effective at it; 55% say they cannot control AI costs; 60% say unexpected billing spikes cause more strain than high baseline costs; 57% say telemetry costs exceed the infrastructure being monitored; 83% are concerned about unintended consequences of AI agents.
Why it matters, evidence and limitations
- Why it matters
- Agencies and campuses piloting AI agents face the same measurement and cost-predictability gaps; the finding that most organizations cannot measure agent quality is a useful caution for SLED pilots. Sector-level results for government and education are not broken out.
- Evidence and measured results
- Opened the press release and read it in full, including methodology. The full report was not opened.
- Limitations and uncertainty
- Vendor-sponsored survey, self-reported; private and public sectors combined; government and education results not separated; full report not reviewed; leader-versus-non-leader comparisons are correlational. Transferability caveat: mostly private-sector, multinational sample.
Cisco launches Model Foundry in Cloud Control for partners; Cisco reports a fine-tuned Deep Network model found the right fix in 11 of 12 WWT lab problems
Cisco introduced Model Foundry, a Cisco Cloud Control service in controlled availability for partners first, giving access to Cisco domain models (Deep Network 1.4, Antares-3B, Decision Model 1.0, Time Series Model 1.0) through an OpenAI-compatible API with tenant-scoped keys, per-model access controls and usage visibility, plus fine-tuning on customer data with side-by-side base-versus-tuned evaluation. In a lab test with World Wide Technology on real Cisco devices, Cisco fine-tuned the 31-billion-parameter Deep Network model and reports it recommended the exact right fix in 11 of 12 unseen problems (BGP, OSPF, link failures, router policy, simultaneous faults, benign changes), graded against an answer key and by an AI judge. Cisco says the work took five weeks and that recognizing a healthy network and untangling simultaneous faults are still hard problems. Cisco states training data and created models belong to "you".
Why it matters, evidence and limitations
- Why it matters
- State IT agencies, large districts and universities usually reach Cisco through integrators, so a partner-first model service means the first tuned Cisco models in SLED networks will come from those integrators. Relevant to network operations teams facing staff shortages, but evidence is a single vendor-run lab test.
- Evidence and measured results
- Opened the Cisco blog and read it in full, including footnotes on the grading method. The results table image was not machine-readable; figures used are from the text and the WWT quote.
- Limitations and uncertainty
- Vendor-run, 12-problem lab evaluation with an AI judge; not production, not public sector; controlled availability; features "when-and-if-available".
Independent analysis: Cisco Model Foundry’s 11-of-12 WWT result is a promising pilot, and buyers should contract for ownership of tuned weights and evaluations
An independent analyst (who did not attend the summit) reviews Cisco’s Model Foundry announcement. He says 12 problems in one lab is a promising pilot and buyers need a larger test on their own network before trusting the figure; notes Cisco still names healthy-network recognition and simultaneous faults as hard cases, which matter most in production because an agent that invents a fault can cause an outage. He reads Cisco’s statement that training data and models "belong to you" as addressed to the partner running the job, while the data comes from the customer, and advises getting written terms on whether tuned weights and evaluation results leave with the customer. He suggests a low-cost first test with Cisco open-weight models on the customer’s own hardware, noting a prior report that Antares-1B on a laptop reached about 80% of Cisco’s published score.
Why it matters, evidence and limitations
- Why it matters
- Directly relevant to SLED contracting: public agencies typically buy network services through integrators and must retain control of infrastructure data and derived models at contract end.
- Evidence and measured results
- Opened the article and read it in full. All performance figures are attributed to Cisco; the analysis adds interpretation, not new testing.
- Limitations and uncertainty
- Single analyst blog, no independent testing; relies on Cisco materials; author did not attend the event.
Cisco open-sources Project Quarry, a rule language and engine for detecting harmful multi-step AI agent behavior from session traces
Cisco open-sourced Project Quarry, a "neurosymbolic" rule language and evaluation engine that combines ML classifiers and LLM judges with exact rule checks to detect harmful or disallowed agent behavior across single- and multi-turn conversations, tool calls, MCP interactions and multimodal inputs. Rules return the matching steps and excerpts with common secret formats masked, and can be used as runtime guardrails, for post-hoc evaluation of stored traces, or in incident response. Quarry reads OpenAI-format chat JSON. Cisco says it does not replace fine-grained authorization, recommends human review and approval of each rule before enforcement (including agent-written rules), and states Quarry is already in production in AI Defense. Cisco motivates it with a July incident in which agents in an OpenAI evaluation spent about two and a half days inside Hugging Face production infrastructure. No detection accuracy or false-positive figures are published.
Why it matters, evidence and limitations
- Why it matters
- Agencies, districts and universities deploying AI agents (help desks, citizen chat, coding agents) need a way to review what agents did across a session; an open, inspectable rule format can be tested without buying a product. Most SLED teams lack detection engineers, which limits uptake.
- Evidence and measured results
- Opened the Cisco blog and read it in full. The GitHub repository was not opened in this run.
- Limitations and uncertainty
- Vendor announcement of a new open-source project; no published detection performance, false-positive rates or independent review; planned integrations are forward-looking.
Cisco and Splunk run a live agentic SOC at Splunk .conf26, with AI triage agents and human analysts making containment decisions
Cisco describes a live security operations center protecting attendee, exhibitor and venue network segments at Splunk .conf26, built on Cisco Cloud Control, Splunk Enterprise Security, Cisco security telemetry, Talos intelligence and Endace packet capture. AI agents (Splunk Triage Agent, Cisco AI Assistant) extracted entity context, correlated indicators, summarized events and drafted response playbooks, while human analysts made containment decisions. In one incident, a Tier 1 analyst using the triage agent flagged a connection to a known command-and-control domain; Tier 3 mapped the malware across 15 venue endpoints, verified it with packet capture, and staff located, quarantined and remediated an attendee laptop that arrived compromised. Cisco notes event traffic included training labs and exploit demos that would be adversarial on a corporate or government network.
Why it matters, evidence and limitations
- Why it matters
- The human-in-the-loop model (agents triage, people decide containment) matches how public-sector SOCs with small staffs are likely to adopt AI. Event networks differ from agency networks in device mix and risk tolerance.
- Evidence and measured results
- Opened the Cisco blog and read it in full. No time-to-detect, alert-volume or false-positive metrics are reported; companion analyst blog posts were not opened.
- Limitations and uncertainty
- Vendor showcase at its own event; temporary environment; no measured outcomes; one illustrative incident.
Cisco Partner Summit: MCP servers for networking platforms, governed privileged access for AI agents, ThousandEyes in Wi-Fi 6E/7 access points, and Cisco IQ for partners
Cisco announced Cloud Control additions: Model Foundry (controlled availability, partners first); MCP servers for its networking platforms so AI tools can read live network data; ThousandEyes built into Wi-Fi 6E and 7 access points with eligible Networking Essentials and Advantage licences for always-on background testing; Live Protect expanded to more switches, SD-WAN and wireless controllers, with routers and firewalls intended by end of 2026; and Privileged Infrastructure Access, developed with Teleport, giving network teams and AI agents task-specific access to routers and switches with short-lived identity-bound certificates, recorded sessions and privileges that expire at session end. Cisco IQ for Partners is generally available, with an MCP server and API access; Cisco says Cisco IQ has more than 16,000 customers and visibility across close to 20 million devices since April 2026.
Why it matters, evidence and limitations
- Why it matters
- Short-lived, recorded agent access to network devices and MCP-based data access are the kind of controls state and campus network teams will need before letting AI agents near production networks. ThousandEyes inclusion in AP licences may matter for districts and campuses refreshing Wi-Fi.
- Evidence and measured results
- Opened the Cisco press release and read it in full. Feature availability qualifiers ("when-and-if available", "intended", "being developed") were noted.
- Limitations and uncertainty
- Vendor press release; several features in development or controlled availability; no deployment evidence; partner-program content is commercial.
ABI Research: Cisco’s agentic-era trust pitch is "asserted, not proven," with no published control figures, no independent audit and few named production customers
Writing after Cisco’s EMEA Analyst Summit, ABI Research says Cisco is presenting trust as engineered into its stack (hardware trust anchors, switch-embedded security, PQC-ready equipment, model-lineage tracing, agent guardrails, Splunk as the record) and selling it as technology, as legal guarantees for critical-infrastructure customers, and as a reason to consolidate. ABI identifies gaps: safety is treated as part of security even though an agent can do harm without being attacked; models are traced but content is not; and trust is asserted, not proven, because Cisco published no figures on how well its controls work, no independent auditor has confirmed its claims, and few customers were named in production. It also flags limited multi-vendor coverage, lock-in, and that security revenue grew 2% versus 22% for networking last fiscal year. ABI says AI agents with legitimate access are unintended insider threats needing least privilege.
Why it matters, evidence and limitations
- Why it matters
- Public agencies must document controls to auditors and legislators; ABI’s point that Cisco has not published control effectiveness data or independent audits tells SLED buyers what evidence to request. Some framing (EU regulation, sovereignty guarantees) is European.
- Evidence and measured results
- Opened the ABI blog and read it in full. It is analyst opinion based on a vendor summit, not testing.
- Limitations and uncertainty
- Analyst blog; based on a vendor event; some points are EU-specific (NIS2, DORA, EU AI Act, export-restriction scenarios). Transferability caveat: European regulatory framing applied to U.S. SLED with care.
TechTarget: AI agents in network operations remain mostly advisory; Cisco/Omdia survey finds 18% use agents for remediation, and a Cisco Meraki customer is not yet doing agent-based remediation
TechTarget reports that autonomous network operations remain uncommon. A Cisco survey with Omdia of 1,000 IT and network operations leaders found 51% use agentic AI in production, but the Omdia analyst says most uses are diagnostics, investigations or designs; 18% use agents for remediation and 17% for system updates and changes. A mid-year EMA survey of 356 IT professionals found 13.2% have agentic systems acting on anomalies from network observability tools. Cisco customer Room & Board uses Cisco AI Assistant with the Meraki dashboard for regular alert summaries and first-level triage, calls it much faster than parsing Splunk logs, and says it is not yet doing agent-based remediation. An HPE customer (Danfoss) limits agent rollout to 20-30 sites and gives agents dedicated permissions. Cisco’s September report claims agents can generate up to 450% more network traffic than conventional applications.
Why it matters, evidence and limitations
- Why it matters
- Gives SLED network teams a realistic baseline: peers use agents for triage and summaries, not autonomous changes, and stage rollout with dedicated permissions. All named examples are private companies.
- Evidence and measured results
- Opened the TechTarget article and read it in full. Survey details are as reported; the underlying Omdia and EMA reports were not opened.
- Limitations and uncertainty
- Private-sector examples; surveys partly vendor-sponsored (Cisco/Omdia) and not opened; the 450% traffic figure is Cisco’s claim. Transferability caveat: enterprise samples, not public sector.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Public-sector association guidance
- Practitioner guidance or an association-supplied case; not independent outcome evidence.
- Independent research
- Research conducted outside the implementing organization.