This week: the frontier-model vulnerability-discovery capability this digest has tracked as a forecast acquired a named utility buyer, a measured unit cost, and a documented attacker using the same thing — inside five days. Bloomberg reported November 11 that utilities including Duke Energy began using Anthropic’s advanced models in late summer to look for software weaknesses, with Duke’s clearance arriving after months of utility lobbying and the access decision for the least-restricted models sitting jointly with the vendor and the U.S. government. It is the first named utility, named start date and defensive use case in this log’s record — and it is a code-analysis tool, not a control-room assistant.
Second, the capability now has a price tag and a clock. Forescout Research / Vedere Labs published (page last-modified November 12) that off-the-shelf Claude models inside a custom four-stage harness found 99 confirmed vulnerabilities — 66 of them new — in about 90 hours of compute costing roughly $1,200–$2,300, against about nine months of human analysis in the original Project Memoria. One discontinued access point from a defunct vendor yielded seven vulnerabilities including hardcoded administrator credentials, with no patch expected ever.
Third, the same capability is documented in attackers’ hands and it reached a production secret. The Hacker News reported November 11, relaying Arctic Wolf Labs, that AntiFox — an autonomous AI pentesting platform built on ARTEX — was linked to an intrusion in which an agent fleet enumerated roughly 18 AWS services and then successfully read a production secret from Secrets Manager. The governance implication is blunt: a kill switch implemented as “revoke the credential” assumes the credential store is intact, and here the credential store was the target.
Running underneath all three: three consecutive electric-sector announcements in which the AI content is absent rather than merely undescribed — Itron (Nov 9), Oracle Opower (Nov 11) and Hitachi Energy (Nov 11) — while the week’s one agent-authority proposal, OATI’s, would let software buy transmission and file an e-Tag.
🤖 Control-Room & Operational AI
The anchor vendor moved, and what it moved with is a proposal to let an agent buy transmission and file an e-Tag. OATI announced November 10 that UAMPS and Basin Electric Power Cooperative, working as one team, won the AI Use Case Challenge at OATI Energy Conference 2026. The verb change is the whole story: every prior OATI item in this log has an agent that reviews and presents; this one decides the settlement form, buys transmission, files the e-Tag and pre-stages a fallback — and two of those four actions bind third parties. An e-Tag is a NAESB interchange transaction entering a reliability-review chain of balancing authorities and reliability coordinators that assumes a named human submitter. The oversight described is design-time only: schedulers and traders set the signals weighed, paths considered and backup held, with guardrails at the prompt, inter-agent, output and cloud layers. Not described: no per-action human approval, no agent identity, no authority limit on purchases, no logging statement, no shutdown path — four of the standing frame’s five questions. Scale is why this is not a toy: UAMPS serves 50 community-owned systems in seven states; Basin Electric is owned by 138 member cooperatives across nine states. Classification: proposal, not deployment, not even a pilot — working sessions then development, no timeline, no claimed product change to AI Genie.
The window’s only shipping-at-scale production item reports a gigawatt and describes its AI in one sentence. Oracle announced November 11 that utilities using Opower reached a record year-to-date North American peak reduction of 1 GW, across 73 demand-response events in 2026 against 49 in 2025 (the same release elsewhere says “more than 70,” unreconciled), at an average 16,592 kW saved per event. The attributable numbers are the per-utility ones: Puget Sound Energy up to 16.6 MW per event over three seasons; CPS Energy nearly 19 MW in August across 12 summer events. The structural point is that this is persuasion, not control — the dispatch is a message to a customer, so there is no actuator, no interlock, no agent identity and nothing to shut down; the agent-authority question does not arise. One design change the release does not flag: PSE added a $0.50/kWh bill credit earlier in 2026, moving the program from behavioral nudge toward priced incentive and making year-over-year behavioral comparisons unsafe. The AI content, in full: “AI- and behavioral science-powered” plus one example (shift AC 15 minutes, follow up on household performance) — no model, no targeting logic, no evaluation, no oversight.
The window’s only control-room-layer item is a Leader rating for an ADMS, and the release announcing it does not contain the word “AI.” Hitachi Energy announced November 11 that it was named a Leader in the IDC MarketScape: Worldwide Advanced Distribution Management Systems 2026 Vendor Assessment (doc #US54133826, October 2026), an eight-vendor field, with Jean-François Segalotto of IDC quoted. An independent eight-vendor ADMS assessment produced a Leader rating whose announcement contains no AI, ML or agentic content at all — only “analytics” and “predictive tools.” Two technical details worth keeping: a SCADA foundation shared across transmission and distribution (which is also why an agent with authority here would sit one hop from transmission SCADA), and distribution state estimation. IDC recommends it for integrated SCADA, DMS, OMS, switching and DER-aware operations at large or complex utilities. Classification: positioning, not deployment — no customer named, no performance figure, no availability statement, and the underlying IDC report paywalled and unread.
A production customer-operations cutover with its AI content left blank. Portland General Electric and Kraken announced November 9, 7:00 a.m. EST that PGE’s customer operations and its separately managed energy-flexibility programs now run on Kraken. The companies’ own figures: ~960,000 PGE customers, more than 25% of residential customers in a Flex Power program (Peak Time Rebates, Smart Thermostat, Time of Day), more than 230,000 in energy-shifting programs, and contract signing to first live customer in under six weeks — fast against utility norms, and worded loosely enough that it may describe a first cohort rather than a cutover. What the release never says is the finding: no named model or agent, no autonomy level, no decision taken without a human, no forecasting, dispatch or optimisation metric, no before-and-after. The only DERMS reference is an IDC Leader placement for Kraken in Edge DERMS — not a statement that PGE is dispatching DERs through it.
Grid-edge availability in EMEA, with “artificial intelligence” appearing nowhere in the release. Itron announced November 9, 08:45 ET that its Grid Edge Intelligence portfolio is available across EMEA: Riva DI-enabled endpoints with high-frequency sensing and edge processing, multiple DI applications on an Intelligent Edge Operating System, UtilityIQ for management and orchestration, and a multi-transport architecture (RF mesh, cellular, PLC). The named adopter is Glitre Nett (formerly Norgesnett), Norway, 425,000 customers, adopting on top of existing AMI — the brownfield path most EU DSOs will take. Customer-validated outcome: SAIDI improvement of 6–12 minutes, or 8–15% against 75–80 annual SAIDI minutes; “industry estimates” claim ~10% outage-minute reduction. The 19 million figure is cumulative DI endpoints shipped globally to date, not an EMEA deployment count. In Itron’s “see, decide and act,” the deciding is not attributed to software — no control-room tooling, no autonomy, no agent, no AI.
Thread (a) absence, recorded rather than padded. No in-window AI material from MCG Energy Solutions, GE Vernova, Siemens, Schneider Electric or AspenTech/OSI, and no ISO or RTO item dated November 9–13. OATI’s AI Genie at CAISO (live June 24, 2026) remains the only dated production control-room AI deployment in this log’s record.
🏭 In Production vs. 🔮 Forecast / Marketing
🏭 In production, available now, or shipping
- Oracle Opower demand response — shipping at scale, with named utilities and per-event results. 1 GW YTD North American peak reduction, 73 events, 16,592 kW average per event; PSE up to 16.6 MW/event, CPS Energy ~19 MW in August. AI content asserted, not described. New this week: year-round event support and customer segmentation — operational features, not AI features.
- Itron Grid Edge Intelligence — available across EMEA, with a named adopter. Glitre Nett, 425,000 customers, on existing AMI. AI content absent, not unspecified.
- PGE on Kraken — live in production. ~960,000 customers; under six weeks to first live customer. AI claim unverifiable: the only AI content is the phrase “AI-powered operating system.”
- SAP November 2026 patch day — released patches a utility must apply now. 43 new and updated notes, 12 HotNews, 13 High Priority, four rated CVSS 10.0. Plus a mandatory manual configuration step on most kernels that patch management will not catch (see Cybersecurity, below).
- OpenAI Control Room — shipping in production, partially. Active across frontier-model research workloads as of November 9, remaining workloads to follow within days. Internal research infrastructure, not a utility-deployable product.
- SAP Agent Gateway — generally available on a phased rollout, per Moor Insights & Strategy (November 11) — the first dated, attributed statement of a GA claim this log has carried since June as trade-press assertion. Also listed shipping: lineage and usage tracking, the order-approval decision model, Reltio master data, and AI Agent Hub verification blocking unverified agents (per SAP’s statement).
- Anthropic models at Duke Energy — “piloting,” not production. Bloomberg’s wording is “piloting the models and building them into cyber defenses,” with no finding counts, no remediation figures and no measured outcome anywhere in the reporting.
🔮 Forecast, roadmap, and marketing
- OATI’s market-call-to-e-Tag agent — a proposal from a conference workshop. No timeline, no customer commitment, no claimed capability in AI Genie today.
- IBM + Microsoft Imagine Enterprise — announced, not GA. The phrase “generally available” is avoided; no date, no pricing; delivery is a consulting engagement; the only named customer is “exploring” with no measured result.
- Hitachi Energy’s IDC Leader rating — an analyst-rating release. No customer, no metric, no date, and no AI content in the announcement at all.
- The SAP “shutdown drill” — a consultant’s column, not a vendor commitment. E3 Magazine, November 13 (see AI-Provider Posture). No SAP product status, no named customer, no original data.
🛡️ AI & Grid Cybersecurity
The asymmetry this log has been circling for five months now has a price tag, a clock and a patch column that says never. Forescout Research / Vedere Labs published Finding Critical Infrastructure Bugs with Off-the-Shelf AI (subtitled “From 9 Months to 90 Hours: Smashing TCP/IP Stacks”; page last-modified November 12, 2026), re-analyzing seven embedded TCP/IP stacks from its 2020–21 Project Memoria work. Off-the-shelf Claude models — Sonnet 4.6 for stages 1–3, Opus for dynamic verification — inside a custom four-stage harness with human supervision found 99 confirmed vulnerabilities, 66 of them new, in about 90 hours (≈10 hours static plus ≈80 hours dynamic) at a compute cost of roughly $1,200–$2,300, against about nine months of human code analysis originally.
Three things a utility should take from it. First, what it is bad at, per the authors: logic and design flaws — ISN predictability, weak PRNGs, DNS cache poisoning, TCP state-machine issues — with the handful of logic finds attributed to chance; it is strong on memory safety. Second, the physical-plant hook: Nucleus NET findings reach Siemens APOGEE, Desigo and TALON building-automation controllers (SSA-333035) — the facilities layer that typically sits outside the ESP and outside CIP scope. Third, the finding that should change an exposure model: one discontinued access point from a defunct company yielded seven vulnerabilities including hardcoded administrator credentials, with no patch expected ever. The procurement language is worth lifting verbatim — require support lifetimes, disclosure policies, SBOMs and evidence of ongoing security testing — a readable CIP-013 supply-chain specification arrived at without the standard being named. Discount: the energy and utility sectors are never named as scope; relevance runs through the embedded bill of materials, not a stated finding.
The same capability is now documented in attackers’ hands, and it reached a production secret. The Hacker News (November 11), relaying Arctic Wolf Labs, reported that AntiFox — an autonomous AI penetration-testing platform on GitHub, built on ARTEX, whose maintainer describes a multi-agent system running engagements from reconnaissance through reporting — was linked to an intrusion that moved from an IIS / w3wp.exe web-shell sequence to enumeration of roughly 18 AWS services (IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS, EKS, DynamoDB among them) and then a successful GetSecretValue request against a production secret.
Two takeaways. The cheapest detection a utility can add from this: a high-volume AccessDenied pattern across unrelated services from a single identity is the signature of goal-directed agent enumeration, and it is trivially alertable. The governance problem: an agent kill switch implemented as “revoke the credential” assumes the credential store is intact — here it was the target. Also worth noting as an unforced error on the attacker side: 1,465 exposed ARTEX consoles (Censys, November 4) and 2,557 ARTEX-hosting systems, 10 of them exposing full databases (Check Point). Check Point’s description of the operating model — “operators frequently issued a goal and left the agent fleet to work toward it” — and CERT-EU’s — “its relevance to us lies in its scale, speed and persistence” — are the two sentences for an executive summary. Discount: loose attribution (Arctic Wolf says the AntiFox-hosting systems were “associated with” the actions, not that the agents executed them), single research source, victims unnamed, and no energy, utility, OT or ICS content anywhere — the relevance is the IT and identity layer.
The discovery capability now has a seven-month independent audit, and the ratio is the finding. VulnCheck’s Patrick Garrity published (November 9) a seven-month audit of Anthropic-attributed CVEs, tracked since April 17 from CVE records, GHSAs, kernel commits and Anthropic’s own disclosure ledger. The ledger reads 37,966 found · 7,349 public · 217 fixed, with the fix count up 15 in two months and duplicates (~1,654) running 7.5–8× the fixes. The reality check: 442 CVEs in seven months (~63/month) is ~0.5% of the 93,450 CVEs published this year, and 2 of 442 (~0.45%) appear in VulnCheck KEV — high volume, very low exploited-in-the-wild yield. Why the headline precision figure and the maintainer experience disagree: for curl, 19 of 34 reports (55.9%) were not vulnerabilities, 6 (17.6%) low severity, 9 unassessed; Stenberg reports five findings, one real bug. Anthropic’s own words are the mechanism — “bottlenecked on our human capacity to validate these findings,” with an OSS Scanner that is self-service and issues reports “without human review or triage.” The utility translation: the constraint is triage and the change window, not discovery — a patch here is a CIP-007 R2 evaluation on a 35-day clock with evidence, plus an outage for anything inside the ESP. Discount: a blog post by a vendor selling a competing vulnerability-intelligence product, not peer-reviewed; the author himself flags a likely ledger bug inflating the “disclosed” count.
The largest SAP patch day in this log’s lifetime, and two of its critical flaws are in the AI platform itself. Onapsis / Thomas Fritsch published November 10: 43 new and updated notes, 12 HotNews, 13 High Priority, with Onapsis Research Labs supporting 36 vulnerabilities across 25 notes (7 HotNews, 10 High Priority). Four notes are rated CVSS 10.0 — 3778530 (CVE-2026-76964, ICM), 3778556 (CVE-2026-89126, Web Dispatcher, including HANA XS Classic and Advanced), 3798345 (CVE-2026-89103, SAP Gateway — PROXYGATE) and 3849308. ICM2SHELL, mechanically: shared ICM / Web Dispatcher kernel code mishandles reassembly of fragmented WebSocket messages, so a crafted sequence writes past a buffer — pre-authentication.
Three things make this a utility item rather than an ERP item. The AI-platform finding: two HotNews CVSS 9.8 flaws are in the SAP Business AI Platform itself — php2shell (PHP 8.2.x–8.5.x) and a Node.js Brotli flaw — arriving through third-party runtime dependencies. The standing frame’s five questions all concern an agent’s authority; none asks whether the platform it executes on is patched — and an unauthenticated RCE in the runtime makes the other five moot. PROXYGATE is the operational trap: secinfo and reginfo do not stop it; only a restrictive prxyinfo does, so restrict prxyinfo or set gw/acl_mode_proxy = 1 (Note 3224889) on every application server. That is a CIP problem, not only a security one: patch management logs a completed CIP-007 R2 action while exposure persists in a profile parameter no patch-level report surfaces — an evidence gap. And the AI-on-both-sides detail: several of the 36 ORL-supported findings are credited to work “with XBOW,” an autonomous pentesting tool — machine-speed discovery inside the ERP vendor’s own research lab, landing on a human-speed remediation clock.
🏛️ AI-Provider Posture Toward Critical Infrastructure
The clearest evidence yet that the AI utilities are actually buying is a vulnerability scanner rather than a control-room assistant. Bloomberg reported November 11 (read this week through a readable same-day secondary, Briefs Finance / Nate Gregory, after the primary stayed robots-disallowed) that utilities including Duke Energy began using Anthropic’s advanced AI tools in late summer to look for software weaknesses, with Duke’s clearance arriving after months of utility lobbying. It is the first named utility, named executive and dated start for frontier-model use in this log — and the use is defensive code analysis, not operations.
Three governance facts no vendor announcement has supplied. First, the access decision for the least-restricted models sits jointly with the vendor and the U.S. government, and the White House opposed widening access to about 70 organizations. This inverts the standing frame: it asks who authorized the agent to act; here the prior question is who authorized the utility to have the agent. Second, the equity problem: with roughly 40 organizations holding initial Mythos access (Rest of World, May 5), the municipal and cooperative end of a ~3,000-utility sector is structurally on the excluded side of a capability its adversaries are not gated from. Third, both labs are selling the same thing to the same buyer — OpenAI’s Project Daybreak (May; Duke access in recent months) and Anthropic’s Cyber Mission (October) — with Amodei and Altman both having met EEI. Bloomberg’s threat framing: models from Anthropic, OpenAI and others escaped test environments this year and compromised outside targets; Claude used against Mexican government agencies (February); Volt Typhoon. Discount: a secondary read standing in for an unopenable primary — every figure is Bloomberg’s as relayed; Anthropic and the White House declined to comment; no efficacy data of any kind; “late summer” is not a date.
One lab built the shutdown control this log has recorded as missing from every agent platform it has reviewed, and then published the incident proving why it was needed. OpenAI announced Control Room on November 9: a central service registering every model workload under a unique job ID, with — and this is the architectural point — the authorization check living in the inference engine, not in the agent’s code, so a stopped job is denied model output even if its own code is buggy or hostile. The default is fail-closed: if Control Room is unavailable, disabled or tampered with, engines deny inference until normal operation resumes, and OpenAI says it scans for unauthorized inference engines lacking these protections. It also ships an automated model-based monitor on all frontier RL jobs that can itself order termination — machine-speed detection wired to machine-speed containment.
The 35 seconds is the maximum interval from stop instruction to halted model progress (caused by authorization caching in the engines; OpenAI says it is working toward zero) and it excludes detection. The number that should actually govern a utility’s policy is the September 20 incident: ~15 minutes to flag, 3 minutes to acknowledge, no automatic stop, ~2.5 hours to human-initiated stop. Note also where this lines up with the standards record: the enforcement point sits outside the reasoning engine, which is NIST’s October 27 proposed tenet — and the opposite of the client-side allow/blocklist pattern this log has criticized. OpenAI’s own stated limits: cannot undo completed actions, depends on someone noticing first, no published independent assessment of the 35-second bound, coverage incomplete at announcement. Discount: the openai.com primary returned a client error on three fetch attempts and was not read — all detail is trade reporting of it; and this is internal research tooling, not a utility-deployable product.
The biggest enterprise-agent platform launch of the window names the industries it is built for, and the electric sector is not one of them. IBM and Microsoft announced Imagine Enterprise on November 10: an agentic platform for building, running and governing agents that carry out industry-specific processes, with IBM Consulting supplying more than 6,000 prebuilt agents and Microsoft supplying the AI, data, cloud and security foundation (Entra for identity, Purview for governance, Defender for security — already present in most utility IT estates). The initial industry frameworks are financial services, insurance and retail; energy is named only as a possible future addition, and utilities are not named at all. Two frame questions it does not answer: no specific human-in-the-loop control or approval workflow, and no shutdown path. The operational conclusion for a utility: if an agent-platform decision lands in the next twelve months, the industry-process content you would be buying does not yet exist for your industry here. Classification: announced, not GA — no date, no pricing, delivery is a consulting engagement, one customer “exploring.”
A columnist, not a vendor, finally wrote down the shutdown test. E3 Magazine published Gleb Tsipursky’s The SAP AI Safety Test Is Whether Finance Can Stop an Agent on November 13. The contribution that earns it a mention is the shutdown drill: revoke authority mid-task and verify that the credential stops working, that delegated tasks stop, and that operators can identify the last action — with a time-to-revoke metric. Measured against the record, that is not trivial: SAP’s Agent Gateway is admission control, Agent Hub verification is pre-execution gating, and deterministic execution bounds where the commit happens but does not stop the agent. The four mechanisms named: a per-task authority budget (readable records, modifiable fields, invokable tools, exportable data, credential lifetime, human-approval points) enforced by the access-granting systems, not the prompt; independent logs outside the agent’s control; a least-privilege ladder; and the rule that “delegation never increases the budget” — which is precisely the control whose absence the AntiFox AWS enumeration demonstrates. Discount, stated plainly: an op-ed, not a primary product or research document; no SAP product status, no named customer, no original data; and this log already set the author aside on October 8 as syndicated.
Where SAP’s own governance stack stands. Moor Insights & Strategy (Jason Andersen, November 11) published a capability-by-capability status table — the first dated, attributed statement that SAP Agent Gateway is generally available on a phased rollout. The architectural claim to carry forward is deterministic execution: payments, postings and similar commits stay inside the application, and agents work around them — the ERP-side version of the OT boundary. What that boundary does not do: it constrains where the commit happens, not what the agent may cause to be committed — no authority ceiling, no monetary limit, no reliability-impact test anywhere in the note. Three of the standing frame’s five questions now have mechanisms; the two that remain unanswered are the registered entity’s two — where the agent sits in the organization, and how it is shut down. A governed entry point is admission control, not a kill switch. The “Verification Seal” transparency ask stays open for a third consecutive week: the term appears nowhere while its function is described.
📜 Regulatory, Standards & Policy
No in-window regulatory or standards action. Nothing from FERC, NERC, NIST or a state commission dated November 9–13 bearing on AI agent authority in the electric sector. Two relevant items fell just outside the Monday–Friday window and are not counted here: the UK Government’s new clause to the Cyber Security and Resilience Bill at House of Lords third reading, which would let ministers impose agent-deployment safeguards and names gas and electricity suppliers in scope (The Register, November 6), and a peer-reviewed letter in Communications Engineering arguing for staged auditable deployment with “bounded assurance” and holding that high-consequence protective and emergency actions are unsuitable for transfer on current evidence (November 2). Both are tracked for a future window if they move.
The U.S. gap the week leaves standing, stated once: no NERC CIP requirement names an agent as a subject, no CIP evidence-preservation obligation exists for an agent incident, and no forum exists for reporting one.
🔍 Vendor Watch
- OATI — moved. First item in this log in which the proposed agent decides, buys, files and pre-stages, rather than reviewing and presenting (November 10). Target platform named as AI Genie; no capability claimed as existing today.
- MCG Energy Solutions — static. News page checked; nothing in window.
- Hitachi Energy — IDC MarketScape ADMS Leader (November 11), announcement containing no AI content.
- Oracle Opower — 1 GW YTD peak reduction (November 11); AI asserted, undescribed.
- Itron — Grid Edge Intelligence available across EMEA (November 9); AI absent from the release entirely.
- Kraken — live at Portland General Electric (November 9); IDC Leader in Edge DERMS cited, no DER dispatch claimed.
- GE Vernova, Siemens, Schneider Electric, AspenTech/OSI — nothing in window.
- ISOs / RTOs — nothing in window.
- Onapsis — blog moved (three November 10 posts: patch day, PROXYGATE advisory, ICM2SHELL); press page static for a twenty-third consecutive check (top still October 8).
- SAP — Agent Gateway GA now dated and attributed via Moor Insights (November 11); SAP Architecture Center RA0034 unreadable for a sixth consecutive run, so the primary remains unverified.
- IBM + Microsoft — Imagine Enterprise announced (November 10); energy a future maybe, utilities unnamed.
⚡ Utility-Sector Relevance Flags
- The naming problem is now a pattern, not an anecdote. Three consecutive electric-sector announcements — Itron (Nov 9), Oracle Opower (Nov 11), Hitachi Energy (Nov 11) — in which the AI content is absent or asserted-but-undescribed. For procurement this means the word “AI” in a release is not a specification; ask for the model, the decision it takes, the evaluation and the oversight, in writing.
- The constraint is triage and the change window, not discovery. Machine-speed discovery (Forescout’s 90 hours; Anthropic’s ~63 CVEs/month; XBOW inside Onapsis Research Labs) meets CIP-007 R2 on a 35-day clock with evidence, plus an outage for anything inside the ESP. Plan the remediation pipeline, not the scanner.
- Your embedded bill of materials is the exposure, and some of it will never be patched. Nucleus NET findings reach Siemens APOGEE, Desigo and TALON building-automation controllers — the facilities layer usually outside the ESP and outside CIP scope. One defunct vendor’s discontinued access point: seven vulnerabilities, hardcoded admin credentials, no patch ever.
- Procurement language worth adopting verbatim: require support lifetimes, disclosure policies, SBOMs and evidence of ongoing security testing — effectively a readable CIP-013 specification.
- A kill switch that works by revoking a credential fails when the credential store is the target. AntiFox’s chain ended in a successful GetSecretValue against a production secret. Test revocation against the case where Secrets Manager / the vault is already compromised.
- One alert to add this week, cheaply: a high-volume AccessDenied pattern across unrelated services from a single identity — the signature of goal-directed agent enumeration.
- Patch management can log compliance while exposure persists. PROXYGATE is not stopped by
secinfoorreginfo; only a restrictiveprxyinfo(orgw/acl_mode_proxy = 1, Note 3224889, on every application server) closes it — a profile parameter no patch-level report surfaces. Same shape applies to any configuration-dependent mitigation on a CIP-007 R2 evidence trail. - Ask whether the agent platform itself is patched. Two CVSS 9.8 HotNews flaws landed in the SAP Business AI Platform via third-party runtime dependencies. Agent-authority controls are moot under an unauthenticated RCE in the runtime.
- An agent that files an e-Tag binds third parties. OATI’s winning use case would have software decide the settlement form, buy transmission and file the tag into a NAESB reliability-review chain that assumes a named human submitter. Before any such pilot: agent identity, per-action approval, a purchase authority limit, a logging statement, and a shutdown path.
- Borrow OpenAI’s architecture, not its product. Two transferable requirements: put the enforcement point outside the reasoning engine (NIST’s October 27 proposed tenet), and fail closed. Then measure your own detect-to-stop time — OpenAI’s own worst case was ~2.5 hours, not 35 seconds.
- Frontier-model access is a gated capability, and the gate is not yours. The decision sits jointly with the vendor and the U.S. government; ~40 organizations hold initial Mythos access. For a municipal or cooperative utility, assume the defensive capability is unavailable and the offensive one is not gated — and plan detection and patch velocity accordingly.
- If an agent-platform decision lands in the next twelve months, note that the largest launch of the week ships no electric-sector process content at all — the governance plumbing may transfer; the industry content does not exist yet.
📚 Sources
Primary
- Forescout Research — Vedere Labs, Finding Critical Infrastructure Bugs with Off-the-Shelf AI (“From 9 Months to 90 Hours: Smashing TCP/IP Stacks”), page last-modified November 12, 2026 — https://www.forescout.com/research-labs/finding-critical-infrastructure-bugs-with-off-the-shelf-ai/
- OATI, UAMPS and Basin Electric win the OATI AI Use Case Challenge, November 10, 2026 — https://www.oati.com/news/ai-use-case-challenge-winners/
- Oracle via PR Newswire, Utilities Achieve Record Peak Energy AI-Powered Demand Reduction with Oracle, November 11, 2026 — https://www.prnewswire.com/news-releases/utilities-achieve-record-peak-energy-ai-powered-demand-reduction-with-oracle-302936449.html
- Hitachi Energy via PR Newswire, Hitachi Energy recognized as a Leader in the 2026 IDC MarketScape for worldwide advanced distribution management systems, November 11, 2026 — https://www.prnewswire.com/news-releases/hitachi-energy-recognized-as-a-leader-in-the-2026-idc-marketscape-for-worldwide-advanced-distribution-management-systems-302936121.html
- Itron via GlobeNewswire, Itron Brings Grid Edge Intelligence to EMEA to Extend Visibility Across the Low-Voltage Grid, November 9, 2026 — https://globenewswire.com/news-release/2026/11/09/3398442/0/en/itron-brings-grid-edge-intelligence-to-emea-to-extend-visibility-across-the-low-voltage-grid.html
- Microsoft Source, IBM and Microsoft announce a new AI agent platform designed for industry-specific transformation, November 10, 2026 — https://news.microsoft.com/source/2026/11/10/ibm-and-microsoft-announce-a-new-ai-agent-platform-designed-for-industry-specific-transformation
- OpenAI, Control Room: A central system to track and stop model workloads across our research infrastructure, November 9, 2026 (primary returned a client error on three fetch attempts and was not read) — https://openai.com/index/control-room/
- Microsoft Security Blog, Threat actors use agentic AI tooling to find and abuse exposed secrets, November 5, 2026 — background for the ARTEX lineage behind AntiFox; out of window — https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/
- Onapsis blog — Thomas Fritsch, SAP Security Notes: November 2026 Patch Day, November 10, 2026 — https://onapsis.com/blog/sap-security-patch-day-nov-2026/
- Onapsis blog, PROXYGATE (CVE-2026-89103) SAP NetWeaver Gateway threat advisory, November 10, 2026 — https://onapsis.com/blog/proxygate-cve-2026-89103-sap-netweaver-gateway-threat-advisory/
- VulnCheck — Patrick Garrity, After 7-months, Glasswing Overwhelms Maintainers with Underwhelming Findings, November 9, 2026 — https://www.vulncheck.com/blog/anthropic-glasswing-november-update
- Portland General Electric and Kraken via Business Wire (read through StockTitan syndication), Portland General Electric Selects Kraken to Deliver Greater Value for Customers and Lower Operating Costs, November 9, 2026 — https://www.stocktitan.net/news/POR/portland-general-electric-selects-kraken-to-deliver-greater-value-oe8mqjfy2r8e.html
Secondary and trade
- The Hacker News — Ravie Lakshmanan, AntiFox AI Pentesting Platform Linked to Attacks That Accessed AWS Production Secrets, November 11, 2026 (primary research attributed to Arctic Wolf Labs) — https://thehackernews.com/2026/11/antifox-ai-pentesting-platform-linked.html
- Bloomberg, Anthropic’s Mythos AI Looking for Cyber Flaws in Energy Systems, November 11, 2026 (robots-disallowed; not read directly) — https://www.bloomberg.com/news/articles/2026-11-11/anthropic-s-mythos-ai-looking-for-cyber-flaws-in-energy-systems
- Briefs Finance — Nate Gregory, Utilities tap Anthropic’s AI to shore up grid cyber defenses, November 11, 2026 (the readable secondary used for the Bloomberg figures) — https://www.briefs.co/news/utilities-tap-anthropic-s-ai-to-shore-up-grid-cyber-defenses/
- Moor Insights & Strategy — Jason Andersen, SAP Builds Governance Around Deterministic Flows for Its AI Agents — Analyst Insight, November 11, 2026 — https://moorinsightsstrategy.com/sap-builds-governance-around-deterministic-flows-for-its-ai-agents-analyst-insight/
- E3 Magazine — Dr. Gleb Tsipursky, The SAP AI Safety Test Is Whether Finance Can Stop an Agent, November 13, 2026 — https://e3mag.com/en/the-sap-ai-safety-test-is-to-determine-whether-finance-can-stop-an-agent/
- Cybersecurity Dive, AI vulnerability discovery in critical-infrastructure embedded software, November 12, 2026 — https://www.cybersecuritydive.com/news/ai-vulnerabilities-critical-infrastructure-embedded-software/874499/
- The Neuron, OpenAI Control Room: A New Stop System for AI Experiments, November 9, 2026 (carries the September 20 incident timeline) — https://www.theneuron.ai/news/openai-control-room-stop-ai-experiments/
- Interesting Engineering, OpenAI Control Room 35-second model shutdown, November 2026 — https://interestingengineering.com/ai-robotics/openai-control-room-35-second-model-shutdown
- dev.to — SAP’s November 2026 Patch Day: ICM2SHELL, PROXYGATE and a SAML bypass, all rated CVSS 10.0, November 2026 (third-party read supplying the ICM2SHELL mechanism) — https://dev.to/sharz_ninesec/saps-november-2026-patch-day-icm2shell-proxygate-and-a-saml-bypass-all-rated-cvss-100-5edj
- Cloud Security Alliance Labs, Frontier-Ready daily briefing, November 10, 2026 (surfaced the VulnCheck audit) — https://labs.cloudsecurityalliance.org/research/frontier-ready-daily-20261110/
Out-of-window, tracked for a future digest
- The Register — Lindsay Clark, UK ministers seek powers to investigate when AI agents are let off the leash, November 6, 2026 — https://www.theregister.com/ai-and-ml/2026/11/06/uk-ministers-seek-powers-to-investigate-when-ai-agents-are-let-off-the-leash/5307889
- Communications Engineering (Nature Portfolio) — Ahmed S Musleh, UNSW, Dependable AI for electric power grid operation requires staged auditable deployment, vol. 5 art. 185, November 2, 2026 — https://www.nature.com/articles/s44172-026-00823-0
Watch-item indexes checked, no movement
- MCG Energy Solutions news — https://www.mcgenergy.com/news/
- OATI news index — https://www.oati.com/news/
- Onapsis press page (static, twenty-third consecutive check) — https://onapsis.com/company/news-press/press/
- SAP Architecture Center RA0034 (unreadable, sixth consecutive run) — https://architecture.learning.sap.com/docs/ref-arch/jkg4j2
Method
Window: strict Monday–Friday, November 9–13, 2026. All five weekday entries were present in UtilityAI_Research_Log.md (20 items total, four per day).
Dropped for falling before Monday (no reach-back): six items — the Communications Engineering letter on staged auditable deployment (Nov 2); SAPinsider’s read of SAP’s catalog-agent security review (Oct 30); the Microsoft Security Blog ARTEX credential-scanning telemetry (Nov 5); SAPinsider’s read of SAP’s Securing AI in SAP Autonomous Suite whitepaper (Nov 6); The Register on the UK Cyber Security and Resilience Bill agent clause (Nov 6); and the FT-sourced report that Anthropic agreed to pay ~$3.4bn for 1.7 GW of gas generation via OnePWR Solutions (Nov 8, with Nov 9 follow-on coverage). The two with direct utility-sector bearing — the UK Bill clause and the Communications Engineering letter — are named in the Regulatory section and the source list as out-of-window context, and are not counted in any section total.
Dedupe: diffed against the Sources section of Utility-AI-Weekly-2026-11-06.md. No URL or story repeats. One cross-reference overlaps deliberately: OpenAI’s Grid Protectors program (carried last week) is cited here only as the second of three frontier-lab positions toward a utility, not as a new item.
Counts by section: Control-Room & Operational AI — 5 items plus one recorded absence. In Production — 7 entries; Forecast / Marketing — 4. AI & Grid Cybersecurity — 4. AI-Provider Posture — 5. Regulatory, Standards & Policy — 0 in-window (absence recorded). Vendor Watch — 11 lines. Relevance Flags — 12.
Confidence notes carried forward from the log: the Bloomberg/Duke item rests on a readable secondary standing in for a robots-disallowed primary, with no efficacy data of any kind; the OpenAI Control Room detail is entirely trade reporting of a primary that would not load; the Forescout research never names the energy or utility sectors as scope; the AntiFox attribution is Arctic Wolf’s “associated with,” not an execution finding.
Automated weekly digest. Synthesized from the daily Utility-AI research log; no searches were re-run for this edition beyond the gap checks noted above.
