InfrastructureIssue 026 records
Utility Management
Controls dominate tonight's utility AI evidence. FERC's FY2026 CIP audit lessons name AI tools as an access-control risk inside grid cyber environments and say to inventory them as service accounts, monitor their outbound paths and assess vendor AI under supply-chain rules. A Frenos vendor benchmark finds top models can assemble OT attack plans from supplied context but inconsistently, and could not attack on their own. Utility Dive's sources say few utilities document AI risk processes. On the deployment side, Florida's 32 FMPA members gained joint-purchased AI pole inspection now being used in the Hurricane Isaias response, with vendor-only benefit figures. AWWA tells water communicators to treat LLM output as a draft. A Tanzanian gas pipeline study models careful leak-detection validation. Gas, metering and stormwater evidence remains thin.
- Evidence records
- 6
- Cross-source patterns
- 2
- Evidence classes
- 2 independent reporting1 government audit1 vendor claim1 public-sector association guidance1 academic research
- Outcomes
- 3 emerging2 cautionary1 mixed
- Source freshness
- 4 new this fortnight2 recent
- Research completed
- 2026-10-31
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
AI tools are now an access-control problem in utility OT
FERC auditors found AI assistants spreading inside CIP-networked environments, with entities handling their access in different ways, and told entities to inventory them as service accounts and monitor their outbound paths. Frenos's vendor benchmark found top models can assemble OT attack plans from supplied context but not execute attacks on their own. Utility Dive's sources report few utilities document how they manage AI risk. Together they put the defender's leverage in limiting what AI tools and intruders can reach and read.
Operating questionWhich AI tools, including vendor-embedded features, can read OT documentation or reach control-system networks today, and does each have an owner, an account record and reviewed access?
Supporting evidenceFederal Energy Regulatory Commission, Office of Electric Reliability (with NERC and Regional Entity staff)Frenos (OT security vendor)Utility Dive (Marlene Wilden)
Utility AI performance claims are running ahead of field validation
The FMPA program's 20x coverage and 80% cost figures are vendor statements made during a live hurricane response. The gas leak model's strong results rest mainly on synthetic leaks and one confirmed event. Utility Dive's sources say consistent performance at scale remains limited. AWWA's panel says fluent LLM output must be verified by a human. Each points to operator-defined validation before reliance.
Operating questionBefore relying on an AI tool, what field-verified test (real events, sampled detections, human fact-checks) will the utility define, and who signs off on the result?
Supporting evidenceFlorida Municipal Power Agency members; Noteworthy AI (reported by Renewable Energy World / Factor This)Dar es Salaam Institute of Technology (Aidan, Greyson, Rugumira)Utility Dive (Marlene Wilden)American Water Works Association
Evidence ledger
Full record · every source keeps its link and limitations
2026 Lessons Learned from Commission-Led CIP Reliability Audits (AI 'super intelligence' tool access)
FERC staff's anonymized summary of non-public FY2026 CIP audits names AI tools (which it calls 'super intelligence' tools, e.g. ChatGPT, Gemini, Microsoft 365 Copilot) as one of three key lessons. Auditors saw growing use of chatbots, personal assistants and automated analytical systems inside CIP-networked environments. Entities generally accounted for them and applied controls, but they handled AI access in different ways. Audited entities said vendors are adding AI to products faster than compliance programs adapt, and that a tool's scope may only become clear after it is active on the network. Mitigations: access granted, verified and revoked under CIP-004-7 R4-R6; communications outside the CIP-005 perimeter understood and monitored; AI tools inventoried as generic service accounts under CIP-007; BCSI access limited under CIP-011; vendor AI assessed in CIP-013 supply-chain plans using NIST AI RMF 1.0.
Why it matters, evidence and limitations
- Why it matters
- Applies directly to NERC-registered municipal utilities, joint action agencies and G&T cooperatives. It also offers distribution-only public power and water utilities a regulator-written control baseline for AI tools that reach OT or sensitive system information. It arrives while New York and Connecticut regulators are asking utilities to inventory AI.
- Evidence and measured results
- Government audit lessons-learned report. The underlying audits are non-public and anonymized, and FERC gives no count of entities, systems or findings for the AI observation. Recommendations are voluntary practices mapped to existing mandatory CIP requirements, not new requirements.
- Limitations and uncertainty
- Bulk electric system scope only; FERC's footnote notes the Bulk-Power System excludes local distribution facilities. Audit sample size, entity types and the frequency of AI findings are not disclosed. The 'super intelligence' label is FERC's terminology for ordinary commercial AI assistants, not a capability claim. Published 13 Oct, before this stream's first edition, but not in the archive; included now because trade coverage on 22 Oct and the pending New York inventory filings make it the current federal control reference.
2026 AI & OT Cybersecurity Report: testing AI models on industrial attack planning
Frenos ran 16,980 tests (13,884 decision attempts, 2,520 planning attempts, 576 simulated runs) on anonymized frontier, US open-weight and foreign open-weight models given supplied facts about OT assets, weaknesses and access. Three frontier versions produced passing complete attack plans in 91.9%, 89.5% and 92.9% of attempts. Only 60.0%, 62.9% and 74.3% of their plans passed every repeat. Other frontier versions scored 17.9-31.9%, the best foreign open-weight 36.2% and the best US open-weight 5.2%. Single-decision scores were 6.0-46.8%, against a random baseline of about 39.5%, and in 16.3% of questions the stated goal did not match the scoring rule. Frenos concludes autonomous attacks were not demonstrated. It recommends segmentation, limited remote access and user rights, detection of unusual sessions, and human review of AI output. It also summarizes Dragos's account of AI-assisted targeting of a Mexican water utility's SCADA/IIoT interface that failed to gain OT access.
Why it matters, evidence and limitations
- Why it matters
- Water and power utilities with small security teams are the targets the report describes. It shows that AI planning ability depends heavily on context the attacker already holds, which shifts the defender's priority to limiting reachable access. It complements the archived FBI/EPA PLC alert.
- Evidence and measured results
- Vendor-run benchmark with anonymized models, supplied facts and fixed action sets, scored by Frenos with a simple harness. The report itself flags a scoring/goal mismatch in 16.3% of single-decision questions. Not peer reviewed and no released dataset was observed. Independent press coverage (Axios, 21 Oct) restates the figures but does not independently test them.
- Limitations and uncertainty
- Vendor research from a company selling OT attack-path analysis, with a commercial interest in the conclusions. Model names are withheld, so results cannot be replicated or mapped to tools a utility uses. Supplied-context tests do not measure discovery or execution. The Dragos water-utility case is summarized secondhand. Published 21 Oct, before the stream's first edition, and not archived; included because it is the most recent quantitative attempt to measure AI-assisted OT attack planning against water and power targets.
Utilities turn to AI to manage rising power demand — from AI (auditability and evidence limits)
Executives describe utility AI as mostly niche decision-support tools (load forecasting, vegetation management, document retrieval, anomaly detection). Large language models are far from acting on physical infrastructure, and agentic architectures touching physical assets are not widely deployed. Cited evidence: a Guardrail Technologies review of federal regulatory filings found about 70% of energy and utility companies named AI as a cybersecurity risk, but only 23% of utilities and 9% of energy companies documented a process to manage it. A National Grid Partners survey found nearly 80% of utility leaders had fully deployed at least one AI application for large-load planning, and 87% said outdated regulatory and rate-case frameworks hinder AI returns. Hardest-to-fill roles are domain experts with AI fluency, data engineers and ML engineers. Sources warn that field records often do not match physical assets.
Why it matters, evidence and limitations
- Why it matters
- Gives public power and co-op leaders an outside view of where utility AI is actually deployed, where evidence is thin, and why auditability and data quality limit expansion. This matters as state commissions begin asking for AI inventories.
- Evidence and measured results
- Trade reporting built on interviews and email statements. Quantitative figures come from a governance-software vendor's filing review and a utility venture arm's survey; methods, samples and definitions are not given in the article. Mostly investor-owned and international utility perspectives.
- Limitations and uncertainty
- Secondary reporting; the statistics come from parties with commercial interests (Guardrail sells AI governance software; National Grid Partners invests in AI startups), and their methods are not described. Little public power or co-op specific evidence. Published 22 Oct, before the stream's first edition, and not archived; included because it is the most recent independent synthesis of utility AI evidence limits.
How Florida municipal utilities are using AI to improve storm damage assessment (FMPA joint purchasing of vehicle-mounted AI inspection)
After a competitive RFP for an AI Utility Infrastructure Data Collection System, FMPA made Noteworthy AI's vehicle-mounted camera platform available to its 32 member utilities. Members can buy, rent or outsource collection under standardized terms and group pricing. The platform geolocates poles and supports asset inventory, GIS development, condition analysis, joint-use audits and post-storm damage assessment, with delivery into systems such as Esri ArcGIS Online. In the current Hurricane Isaias response (landfall 9 Oct near Destin), the vendor says field staff can cover up to 20 times more ground per day. The vendor also says program pricing can be up to 80% below conventional collection. At the City of Starke, about eight hours of collection inventoried roughly 2,300 poles, and a third-party audit found uncounted pole attachments that helped offset costs. Members could form a camera-equipped mutual-aid network.
Why it matters, evidence and limitations
- Why it matters
- A concrete joint-action-agency procurement model that lets small municipal utilities buy AI field data on shared terms. It is being used now in a live hurricane response, and it ties to FEMA documentation needs.
- Evidence and measured results
- Trade article relying mainly on vendor executives' statements plus one utility director quote; APPA's 15 Oct article (also inspected) says the pricing and deployment claims come from Noteworthy AI. No independent measurement of the 20x coverage, 80% cost or damage-assessment accuracy figures is reported.
- Limitations and uncertainty
- Benefit figures are vendor claims; the Isaias response is ongoing and no outcome data is reported. FEMA reimbursement benefits are possible, not demonstrated. The Starke result is a single small-city example with no reported accuracy validation.
Large language models are a tool, not a replacement, for utility communications
AWWA summarizes its webinar 'Applying AI in Water Utility Communications' (more than 800 registrants). Panelists say all LLM output must be validated and edited by a human because fluency does not guarantee accuracy, and drafts stay drafts until verified for accuracy and voice. Recommended practices: specific prompts describing situation, task and audience; configuring tools with the utility's brand guide and past communications; and feeding newly published material back so outputs stay aligned. Panelists also describe a 'visibility gap'. AI search answers about a utility may be inaccurate and the utility cannot see who asked or what was answered. They advise controlling the consistency of public online content (generative or answer engine optimization), since AI draws heavily on news, Q&As and earned media.
Why it matters, evidence and limitations
- Why it matters
- Water utility communications, including boil-water notices, rate changes and incident updates, depend on public trust. This is the sector association's current practice guidance for staff already using general-purpose LLMs.
- Evidence and measured results
- Association news article summarizing expert panel advice (a PR consultant, an AI consultant and a university communications professor). No measured outcomes, survey data or evaluation of accuracy are reported.
- Limitations and uncertainty
- Practitioner advice, not evaluated evidence. Panelists include consultants with commercial services. Emergency communications and accessibility requirements (plain language, translation) are not addressed in the article. The full webinar recording was not reviewed.
ULeaD: A Hybrid Isolation Forest-LSTM Autoencoder Model for Unsupervised Gas Pipeline Leak Detection: The Kinyerezi-Ubungo Case Study
An unsupervised model combining Isolation Forest and LSTM autoencoder anomaly scores (weights 0.3/0.7) was trained on twelve months of one-minute SCADA data (2022, chronological 70/15/15 split) with ten expert-screened inputs. Threshold and a three-of-five persistence rule were fixed before testing. On 25 normal test windows the timestamp-level false-alarm rate was 1.73%. In 200 synthetic flow-loss experiments (1-10% of inlet flow), mean precision was 0.89, recall 0.94 and F1 0.88 for 2-10% leaks, with pooled AUC-PR 0.906 and mean detection delay of 10 minutes, all better than either component (p<0.01). On the single confirmed field leak (28 Dec 2022), held out from development, ULeaD alarmed 8 minutes after onset versus 15 and 19 minutes for the components. Authors say field performance needs validation against more confirmed events.
Why it matters, evidence and limitations
- Why it matters
- Limited direct relevance to U.S. public gas systems, which run low-pressure distribution networks rather than short transmission lines. It is included because it shows an evaluation design municipal gas utilities can ask vendors for: leakage-free chronological splits, pre-fixed thresholds, false-alarm rates and held-out real events. It addresses the scarcity of confirmed leak records that also limits AI leak tools in water and gas distribution.
- Evidence and measured results
- Retrospective single-pipeline study using SCADA history, synthetic flow-only leak injections and one confirmed field leak. Statistical comparisons against component models; no comparison with commercial leak-detection systems or API RP 1175 programs. Journal peer-review process not stated on the article page.
- Limitations and uncertainty
- One 12.3 km transmission pipeline, one year of data, synthetic flow-only leaks and a single real leak. Transfer to U.S. municipal distribution systems is unproven. Peer review not confirmed. Published 13 Oct, before the stream's first edition; included because gas was uncovered in that edition and this is the most recent openly accessible gas AI evaluation found.
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.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Government audit
- An oversight review of performance, controls, or operations.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Public-sector association guidance
- Practitioner guidance or an association-supplied case; not independent outcome evidence.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.