---
title: "Gated Dual-Use Models and Adaptation in the Drone War of Attrition"
date: 2026-11-02
lang: en
timezone: Asia/Tokyo
---

# Gated Dual-Use Models and Adaptation in the Drone War of Attrition

A biology model with a looser refusal boundary, 433 drones under cloud cover, and a domestic intrusion where state involvement is now the question.

## Today’s key points

The weekend's developments shifted the centre of gravity from raw capability to the design of use: who gets access and under what conditions. OpenAI opened a same-weights model to vetted biologists, Ukraine and Russia contested the conditions of interception around weather and computer vision, and in Japan the attribution of an intrusion began to connect with the government's institutional decisions.

OpenAI released GPT-Rosalind-6, which shares weights with GPT-6 Astra but refuses fewer dual-use biology queries, to approved organisations only. The Preparedness classifications stay at High for biological and chemical risk and Critical for cybersecurity.

Russia launched 433 drones in a single night and Ukraine's Air Force reported neutralising 371. The mix switched from the jet-heavy raid of the previous night to piston-engined Shaheds under overcast skies, a choice analysts read as targeting the visual acquisition conditions of interceptor drones.

Kyodo reports that GMO Internet Group has told the government that a North Korea-linked group is likely behind the intrusion. The first press conference on Monday afternoon opens the door to a decision on the first use of active cyber defence and to public attribution.

A senior INDOPACOM officer said Japanese partners used the Maven Smart System for the first time and contributed to joint strikes. No Japanese account has yet been found, leaving open whether this was exercise-bound or routine use.

## Main developments

### OpenAI releases GPT-Rosalind-6 under trusted access: same weights, looser biology refusals, cyber still rated Critical

On 30 October OpenAI published the system card addendum for GPT-Rosalind-6, a model for biological research. It uses the same weights as the flagship reasoning model GPT-6 Astra, with no added training data or post-training, and refuses fewer dual-use biology queries. Access is limited to approved organisations, and the model is rated High in biological and chemical capability and Critical in cybersecurity.

#### A refusal boundary that moves without retraining

The addendum states that GPT-Rosalind-6 uses the same weights as GPT-6 Astra and that no training data was added and no additional post-training was performed for the release. The difference between the two models therefore lies in the operational boundary of how far the model answers dual-use biology questions and in the distribution condition of who receives it, rather than in the weights. The model is designed to be more useful for legitimate advanced research while still refusing requests that would meaningfully enable severe biological weaponisation.

The arrangement amounts to a two-tier release of a single capability: conservative refusals for the public, looser ones for vetted researchers. Separating the capability assessment from the vetting of recipients resembles end-user screening in export controls. Holding capability fixed and managing risk through access design has now been set out as one template for handling dual-use frontier AI.

#### The basis for the High rating and the unmet Critical thresholds

In the capability evaluations the model exceeded the indicative High thresholds on three of four evaluations after adjusting for refusals. On the troubleshooting benchmark it answered 491 of 832 items, far above the indicative 36.4%. On biorisk knowledge it scored 78.3% against an 80% threshold, rising to 85.0% after refusal adjustment. Against GPT-Rosalind-5.5, MedChemBench v7 rose from 28.9% to 49.6%, GeneBench Pro v13 from 16.7% to 37.7% and LifeSciBench Gold from 44.0% to 61.3%.

None of the four Critical-level evaluations, covering AAV capsid design, SHP2, coronavirus–ACE2 binding and phage–plasmid prediction, met its indicative threshold, so OpenAI placed the launch at High, below Critical. In cybersecurity the model inherits GPT-6 Astra's Critical rating and retains that model's standard safeguards including real-time blocking. Task success under historical jailbreak attacks was 14.3%, close to Astra's 15.7%.

#### Conditions of use: vetting, least privilege and human review

Vetting checks for a legitimate life-sciences or related mission, a credible use case, appropriate research and operational governance, and the ability to control access safely. Non-government applicants undergo business verification and compliance screening. For government entities a separate trusted-access pathway covers approved biosecurity, biosafety, public-health preparedness and biodefense work. The addendum says the release does not broaden eligibility beyond prior announcements such as the Rosalind Academic Plan.

Approved customers must restrict access to authorised users with a legitimate need, apply least-privilege controls, onboarding and insider-risk controls, and apply qualified human review before outputs inform experiments, publications or consequential decisions. They also need processes to remove access when a user no longer needs it, and OpenAI can restrict or revoke access for misuse. Pricing and customer names do not appear in the addendum; the prices and pharmaceutical partners cited in some coverage remain unconfirmed.

#### Questions for evaluators and regulators

The release shows frontier safety frameworks shifting their weight from whether to train and publish toward the conditions and auditing of distribution. Because the same weights carry a cyber capability rated Critical, the model handed to biologists also carries top-tier cyber capability, so the effectiveness of vetting, insider-risk management and revocation now matter as much as the capability assessment itself. The evaluations rest on internal metrics without published third-party replication, so the scope of verification by the US and UK evaluation bodies is the next question.

#### What to watch next

Whether US CAISI or UK AISI publish independent evaluations; disclosure of government adopters on the trusted-access pathway; the full tables in the addendum PDF, including measured values on the Critical evaluations.

#### Sources

- [S1] [GPT-Rosalind-6 System Card Addendum](https://deploymentsafety.openai.com/gpt-rosalind-6) — OpenAI; 2026-10-30; primary

- [S2] [GPT-6 Rosalind shows OpenAI now sells access, not just models](https://www.i-scoop.eu/gpt-6-rosalind-trusted-access/) — i-SCOOP; 2026-10-31; reporting

- [S3] [GPT-6 Rosalind Research Puts Access Choices at the Center](https://forkast.news/gpt-6-rosalind-research-access-boundaries/) — Forkast; 2026-11-01; reporting

### GMO reported to have told the government a North Korea-linked group is likely behind the intrusion; a fake victim forum surfaces ahead of Monday's first press conference

Kyodo reported on the night of 1 November that GMO Internet Group has told the government that, judging from traces including the characteristics of the malware used, a North Korea-linked hacking group is likely involved in the intrusion. The same day brought reports that a suspicious site soliciting information from victims may be linked to the attackers. Chairman and CEO Masatoshi Kumagai holds the first press conference since the incident on Monday afternoon.

#### What five matching IP addresses do and do not show

According to Kyodo, GMO assessed from multiple traces, including the characteristics of the malware, that a North Korea-linked group is likely involved and reported this to the government. In Kyodo's own analysis with security experts, five IP addresses that connected immediately after the attackers' claim site went up on 27 October matched addresses previously used by North Korea-linked groups. GMO handed substantive information on the attackers to police on 29 October.

The claim site contained traces pointing to several countries including the United States and Japan, which experts read as possible noise inserted to confuse investigators. Kyodo itself notes that matching IP addresses alone cannot establish that the same group was responsible, so attribution currently rests on the analysis of the victim and the press. An official judgement awaits the police and the National Cyber Office.

#### A fake victim forum and a warning from two years ago

NHK and Kyodo reported on 1 November that a site called the “information sharing forum,” which solicited information from victims, had been prepared before GMO disclosed the breach on 21 October. Its source code partially matched the attackers' claim site, and the hosting provider, Japanese-language code, folder names and access settings were common to both. The site has already closed with almost no visitors, and the attackers may have intended to use victims to pressure GMO.

Nikkei reported on 31 October that group company GMO Cybersecurity by Ierae had called for a fundamental review of the systems in April 2024, and Kumagai acknowledged on his blog that the depth and speed of the response had been insufficient. Tips under the bounty scheme reached 428 by 18:00 on 30 October, and applications for condolence payments and compensation close on 6 November.

#### A separate nine-year misconfiguration and the governance questions for the press conference

A separate matter surfaced on 1 November. In GMO Trust Login, an identity-management service of certificate subsidiary GMO GlobalSign, non-public files had been placed in a public storage bucket used to serve logo images and were retrievable from outside from November 2017 until 15 September 2026. The third notice on 30 October widened the scope to 2,492 companies and 45,030 individuals and confirmed retrieval by third parties. This was a misconfiguration distinct from the intrusion, with a final report to the Personal Information Protection Commission due on 12 November and main countermeasures by 30 November.

Monday's press conference will therefore face questions on the intrusion route and attribution in the Desktop Cloud case, on the response to the internal warning two years ago, on group-wide configuration management and on the scale of payments and compensation. Any statement on attribution feeds directly into the policy the government set at the 26 October ministerial meeting on cyber and into its decision on the first request for active cyber defence.

#### What to watch next

GMO's account of attribution, intrusion route and compensation at Monday's press conference; official attribution by police and the National Cyber Office; the 6 November deadline for compensation claims and publication of the government's emergency package.

#### Sources

- [S1] [【独自】ＧＭＯ攻撃、北朝鮮ハッカーか　複数の痕跡、政府に報告](https://www.oanda.jp/lab-education/market_news/kn_2026110101000928/) — 共同通信（OANDA転載）; 2026-11-01; reporting

- [S2] [GMOの被害者情報集める不審サイト　ハッカー側と関連か](https://news.web.nhk/newsweb/na/nd-20261031de69043) — NHK; 2026-11-01; reporting

- [S3] [GMOトラスト・ログイン　2,492社・45,030人分が約9年間外部から取得可能](https://www.oanda.jp/lab-education/market_news/kn_2026110101000863/) — 共同通信（OANDA転載）; 2026-11-01; reporting

- [S4] [GMO被害者向け不審サイトの解説](https://innovatopia.jp/cyber-security/cyber-security-news/119600/) — innovatopia; 2026-11-01; reporting

### Ukraine says Russia launched 433 drones in one night and 371 were neutralised; a switch to piston Shaheds under cloud cover, and a third refinery struck in three days

Ukraine's Air Force said on 1 November that from 18:00 on 31 October Russia launched 433 Shahed-type drones and decoys, of which 371 were shot down or suppressed. The mix returned to piston-engined Shaheds under overcast skies after a raid the previous night in which 72 of 134 drones were jet-powered, a choice analysts read as exploiting the visual acquisition conditions of interceptor drones. The same night Ukrainian drones struck the Kstovo refinery, Russia's fourth largest, the third refinery hit in three days.

#### The composition of the 433 and impacts at 16 sites

According to the Air Force, 64 of the 433 were jet-powered, launched from Oryol, Kursk, Shatalovo, Bryansk, Millerovo and occupied Donetsk. By 07:30, 371 drones including 47 jets had been shot down or suppressed by aircraft, surface-to-air missiles, electronic warfare, unmanned-systems units and mobile fire groups, with impacts at 16 locations and debris at 14. More than 40 drones were still airborne at the time of the statement, and no missiles were reported.

Damage included energy infrastructure in Vinnytsia and Kyiv regions, four injured in Rivne, a child injured in Poltava and five houses in Lviv, and the Energy Ministry said parts of eight regions were without power on Saturday morning. President Zelensky said in his weekly summary that almost 2,000 drones and 44 missiles had been used in the past week. The kill and suppression figures are the Air Force's own and remain unverified independently.

#### Cloud cover against interceptor crews, and an upgraded ZIRKA that tracks by computer vision

Defense Express and electronic-warfare specialist Serhii Beskrestnov point to the piston-heavy composition of this raid. Interceptor-drone pilots rely heavily on visually acquiring the target, and thick cloud degrades the conditions for chasing slow targets, so Russia appears to have used the weather to blunt interceptor effectiveness. This is one analyst's inference and Russian intent is unconfirmed, but it suggests a pattern of alternating jet raids and mass piston raids according to weather.

On the defending side, a unit of the 412th “Nemesis” Brigade of the Unmanned Systems Forces was reported on 1 November to have downed a jet-powered Geran-5 flying at about 400 km/h at roughly 3,500 m with an upgraded ZIRKA interceptor. The aircraft runs a trained computer-vision model on a Raspberry Pi 5 that detects Shahed-type features at about 1 km and tracks autonomously after operator confirmation. Date and place of the intercept were not given, so the claim rests on footage. Coping with cloud and onboard autonomous tracking define the interceptors' next tasks at once.

#### The Kstovo strike and a private company claiming it

In the early hours of 1 November explosions and fire hit the Lukoil NORSI refinery at Kstovo in Nizhny Novgorod region, with a capacity of 15–17 million tonnes a year. Governor Gleb Nikitin described a massive drone attack that significantly damaged an industrial facility and claimed more than 100 drones were downed. Fire Point, maker of the FP-1 long-range drone, claimed the strike without giving numbers or damage. No General Staff confirmation had been found by Saturday evening; the strike itself is corroborated by the governor and multiple monitoring channels, while the extent of damage awaits satellite imagery.

The strike overlaps with the Wall Street Journal's 29 October report of a pause in refinery strikes at the US administration's request. Kyiv officials told the Kyiv Independent that no such order existed and only a brief pause occurred, while Russian Deputy Prime Minister Novak said Ukraine had resumed attacks on three refineries and a gas-chemical complex. Satellite analysis assesses that the 30 October strike on the Novokuibyshevsk refinery in Samara region damaged both primary distillation units and halted all primary processing.

#### About 60 deep strikes in October at a median 551 km

Militarnyi's geospatial tally counts about 60 geolocated deep strikes in October: 20 on oil and gas, 10 on military sites, 9 on defence industry, 9 on infrastructure, 6 on chemicals, 4 on logistics and 2 on power. The median distance from the front was 551.5 km and the maximum 2,508 km at Omsk; at least eight combat aircraft (seven Su-34 and one Su-35) were confirmed destroyed, and Russia acknowledged eleven aircraft lost in October. Neptune was used in three strikes and Storm Shadow at Oryol and Yelets. The run of refinery strikes extends this distribution.

#### What to watch next

Satellite-based damage assessment of Kstovo and any General Staff confirmation; the Russian drone mix and interception rate on the next overcast night; refinery strikes and the US response after the 3 November midterms.

#### Sources

- [S1] [Russia Attacks Ukraine With 433 Drones: Air Defences Neutralise 371](https://en.lb.ua/news/2026/11/01/47526_russia_attacks_ukraine_433.html) — LB.ua; 2026-11-01; reporting

- [S2] [Ukrainian Drones Strike LUKOIL Oil Refinery in Kstovo, Fire Breaks Out](https://militarnyi.com/en/news/ukrainian-drones-strike-lukoil-oil-refinery-in-kstovo-fire-breaks-out/) — Militarnyi; 2026-11-01; reporting

- [S3] [Fire Point confirmed the strike on the Kstovo Oil Refinery](https://unn.ua/en/news/fire-point-confirmed-the-strike-on-the-kstovo-oil-refinery-one-of-russias-largest-oil-refineries) — UNN; 2026-11-01; reporting

- [S4] [Upgraded ZIRKA Interceptor Drone Destroys Geran-5 at 400 km/h](https://militarnyi.com/en/news/upgraded-zirka-interceptor-drone-destroys-geran-5-at-400-km-h/) — Militarnyi; 2026-11-01; reporting

- [S5] [Ukraine war latest: Russian oil refinery in Kstovo ablaze after reported overnight drone strike](https://kyivindependent.com/ukraine-war-latest-russian-oil-refinery-in-kstovo-ablaze-after-reported-overnight-drone-strike/) — Kyiv Independent; 2026-11-01; reporting

- [S6] [Ukrainian Strike Damaged All Primary Refining Units at Novokuybyshevsky Oil Refinery](https://militarnyi.com/en/news/ukraine-damaged-all-units-novokuybyshevsky-oil-refinery/) — Militarnyi; 2026-10-31; reporting

- [S7] [Deep Strikes in October: Missile Production, Oil Refineries, and Russian Aircraft](https://militarnyi.com/en/news/deep-strikes-in-october-missile-production-oil-refineries-and-russian-aircraft/) — Militarnyi; 2026-11-01; reporting

### AP satellite analysis finds Iranian strikes damaged 26 US sites, hitting radars first; cheap mass drones unsettle the premise of fixed bases

Drawing on commercial satellite imagery and four regional officials, the Associated Press reported on 31 October that more than 115 Iranian and proxy drone and missile strikes between the outbreak of war in February and August damaged or destroyed structures at 26 US bases and sites. The strikes went for radars in the first two days, then aircraft and hangars, then housing. The Pentagon responds that forces and bases are well protected.

#### More than 115 strikes in the order of radars, aircraft, buildings

According to AP, radar enclosures at Camp Arifjan in Kuwait and the naval support activity in Bahrain were destroyed within the first two days, and portable radars were hit at Prince Sultan Air Base in Saudi Arabia and Muwaffaq Salti Air Base in Jordan. The Jordan radar was one of only 13 of its type in the US Army at about half a billion dollars each, and the Qatar radar came from a 1.1 billion dollar contract in 2013. N.R. Jenzen-Jones of ARES said Iran prioritised the backbone of US air defence.

In the first month more than 50 strikes at 18 locations targeted aircraft, damaging an E-3 early-warning aircraft, and the Congressional Research Service counts 15 US aircraft damaged. Ali Al Salem Air Base in Kuwait was struck more than 30 times, with at least 15 hangars destroyed. A drone attack on Port Shuaiba on 1 March killed six US soldiers. The Pentagon estimates damage above 5 billion dollars and CSIS puts structural rebuilding alone above 4 billion. CENTCOM declines to detail damage on operational-security grounds.

#### Withheld imagery and the Pentagon's rebuttal

AP notes that major imagery providers with US government contracts blocked or delayed releases at Washington's request, so the gap was filled from other suppliers such as Airbus OneAtlas. Pentagon spokesman Sean Parnell replied that forces and bases are well protected and stronger than ever. Two regional officials said the US has moved many aircraft, air-defence systems and communications equipment out of the Gulf, and the Wall Street Journal reports that relocating bases beyond Iranian range is under consideration. The overall picture of damage thus rests on independent imagery analysis amid information control and repositioning.

#### What cheap mass drone attacks ask of fixed bases

Raphael Cohen of RAND said modern forces are susceptible to large-scale, cheap drone attacks and that the idea of relying on fixed infrastructure is over. The sequence of blinding sensors first, then aircraft, then personnel facilities is a textbook suppression-of-air-defence pattern executed with inexpensive drones. Layered counter-drone defence and the dispersal and hardening of bases therefore bear directly on the design of forward bases, including in the Indo-Pacific. The AP piece does not mention AI and gives no breakdown of the systems used, so the degree of autonomy involved is unverified.

#### What to watch next

Official damage disclosure by CENTCOM or the Pentagon; relocation of bases beyond Iranian range and FY27 funding for counter-drone defence; whether operations against Iran resume after the 3 November midterms and whether Gulf states permit base use.

#### Sources

- [S1] [Iran targeted US air defenses before hitting aircraft and buildings, AP analysis finds](https://abcnews.com/US/wireStory/iran-targeted-us-air-defenses-hit-aircraft-buildings-137733589) — AP（ABC News配信）; 2026-10-31; reporting

- [S2] [Iran blinded US bases before striking military assets, AP analysis finds](https://tribune.com.pk/story/2638020/) — The Express Tribune（AP配信）; 2026-11-01; reporting

- [S3] [Satellite images reveal billions in damage to US bases from Iranian attacks](https://www.theweek.in/news/middle-east/2026/11/01/satellite-images-reveal-billions-in-damage-to-us-bases-from-iranian-attacks.html) — The Week; 2026-11-01; reporting

### Two days before the US midterms, federal–state threat sharing is reported at its weakest since 2016, and chatbots are citing an AI-generated fake local-news network

CNN reported on 1 November, citing about a dozen election-security experts, that the channels for passing foreign-threat intelligence from federal to state and local election officials are at their weakest since 2016. Intelligence assessments say Russia is attacking Democratic candidates with AI-altered videos while China targets anti-China House candidates. In parallel, a network of 37 AI-generated fake local-news sites has been found in chatbot answers, showing how AI has entered the information layer of the election.

#### A shrunken CISA and channels that no longer move

According to CNN, a mid-October classified briefing, the first since 2024, was interrupted by technical failures and rescheduled. A last-minute DHS online threat-sharing platform was described as lacking credibility, and David Becker of the Center for Election Innovation and Research said he knows of no election official actively sharing cybersecurity information with federal authorities this cycle. CISA has lost about a third of its staff, and election-security grants moved from 55 million dollars in 2024 to 15 million in 2025 and 45 million in 2026.

A CISA spokesperson countered that an information-sharing tool is available on request and that the agency remains fully equipped. Scott Leiendecker of Liberty Vote, whose machines are used in 26 states, called federal cyber support disappointing, and Arizona CISO Michael Moore warned that politicising threat-intelligence sharing will make people stop sharing. Unsubstantiated remarks by DHS Secretary Markwayne Mullin about election-system vulnerabilities have further deepened state-level distrust.

#### Russia with AI-altered video, China with targeted candidates

Intelligence assessments say Russia is deploying disinformation to denigrate Democratic candidates, and Graphika and others tracked accounts suspected of being Russian operatives posting AI-altered videos on X, including a fabricated sexual-assault allegation against Texas Democratic Senate nominee James Talarico. Rep. Jim Himes of the House Intelligence Committee said Russia is throwing everything out there, that China is extremely targeted at anti-China House candidates, and that Iran has been relatively quiet in the first major election since the US–Iran war began in February.

#### A 37-site fake local-news network cited by chatbots

In research by the security firm VisionHeight reported by NPR on 30 October, 37 sites posing as local newspapers, blogs and state or national outlets carried content promoting ten mostly Democratic House candidates. Thirty-six of the 37 domains were registered within eight minutes on 15 September, about 100,000 words of templated text were posted on 30 September, and the phrase “the stronger choice” appeared more than 360 times. All sites allowed 17 AI crawlers and carried llms.txt, and OpenAI's chatbot cited the network in 19 of NPR's 37 queries. The campaigns and the DCCC deny involvement, and the operator is unidentified.

In a separate Washington Post test asking six chatbots to build voter guides for an address in every county, the bots failed to name the House district 48% of the time and were wrong 22% of the time when they did, and Microsoft's Copilot cited The Ballot Brief, run by a former Rumble executive, in 75% of answers. Right-leaning outlets branded the 37 sites a Democratic scheme, though NPR did not establish the operator. The weakening of official information sharing and the contamination of an AI-mediated information layer are advancing together, which marks this election.

#### What to watch next

Election day on 3 November and official CISA/ODNI assessments of foreign interference; identification of the operator of the fake local-news network; post-election reports from OpenAI's election operations centre and other companies on handling voting information.

#### Sources

- [S1] [Election officials brace for foreign threats and an unpredictable Washington](https://localnews8.com/politics/cnn-us-politics/2026/11/01/election-officials-brace-for-foreign-threats-and-an-unpredictable-washington/) — CNN（LocalNews8配信）; 2026-11-01; reporting

- [S2] [Pro-Dem phony news sites built to fool AI chatbots](https://yahoo.com/news/politics/articles/pro-dem-phony-news-sites-213028540.html) — New York Post（Yahoo配信）; 2026-10-30; reporting

- [S3] [AI voter guides riddled with errors](https://jewishworldreview.com/1026/ai-voter-guides-errors.php) — The Washington Post（Jewish World Review配信）; 2026-10-29; reporting

- [S4] [Foreign Threats to US Elections Are Maturing as US Defenses Erode](https://www.justsecurity.org/161370/foreign-threats-us-elections-maturing/) — Just Security; 2026-10-20; reporting

### INDOPACOM says Japan used the Maven Smart System for the first time and contributed to joint strikes, as the US, Japan and Australia share a single common operational picture

Brig. Gen. Michael Smith, INDOPACOM's director for C4 and cyber, told TechNet Indo-Pacific in Honolulu on 29 October that 2026 was the first year in which US forces and allies ran warfighting applications multilaterally on a single common operational picture, and that Japanese partners used the Maven Smart System for the first time, contributing to joint strikes and closing kill chains. The backbone is the PACOM Mission Network; at Yama Sakura 91 in September the US, Japan and Australia processed 200 digital fire missions in a day.

#### First multilateral operation on the PACOM Mission Network

According to Smith, the PMN is a zero-trust network spanning classification levels that has widened data sharing from bilateral to multilateral. Deputy commander Lt. Gen. George Rowell said that over the past year 17 exercises were run in nine host nations with 35 countries, using common tools, common data flows and language translation for the first time, and that assured command and control is inseparable from the concept of deterrence in the Pacific. The use of Maven was presented as a case of allies sharing AI-enabled target processing on this common backbone.

#### Two hundred digital fire missions a day at Yama Sakura 91

In the account of Gen. Ronald Clark of US Army Pacific, reported by AFCEA's SIGNAL on 27 October, the nine-day Yama Sakura 91 that ended on 10 September saw the PMN enable real-time trilateral fires among the US, Japan and Australia, with 200 digital fire missions processed on day six. Maven was scaled on day five, and the 25th Infantry Division printed 45 drones a week. Japan's use can therefore be read as already embedded in the fire-coordination workflow, at least in an exercise setting.

#### A placement in Japanese operations that awaits Tokyo's account

No Japanese-language coverage or Ministry of Defense statement on these remarks had been found by 1 November. Reports in August described the ministry as still considering Maven for command and control. Whether this first use was an exercise-bound trial or the start of routine use in joint fire coordination will therefore be settled only by Japan's own account. Sharing AI-enabled target processing on an allied common backbone touches on information security, rules of engagement and where decisions on the use of force sit under domestic law, making it a likely subject for parliamentary explanation.

#### What to watch next

Any explanation from the Ministry of Defense or Joint Staff; references to the government segment and Maven in Palantir's 2 November earnings; disclosure of PMN and Maven use at the next US–Japan–Australia exercise.

#### Sources

- [S1] [Bilateral no more: PACOM officials highlight multi-national data sharing network](https://breakingdefense.com/2026/10/bilateral-no-more-pacom-officials-highlight-multi-national-data-sharing-network/) — Breaking Defense; 2026-10-30; reporting

- [S2] [USARPAC, Allies Deploy Long-Awaited PACOM Mission Network](https://www.afcea.org/signal-media/usarpac-allies-deploy-long-awaited-pacom-mission-network) — AFCEA SIGNAL; 2026-10-27; reporting

### China-nexus actors reported to have rooted 25 PAN-OS servers using an AI-built scanning platform, with victims in Japan, Taiwan and South Korea and a US government-affiliated environment

On 29 October the security firm Dream reported that attackers assessed as China-linked used an AI-enabled scanning platform called “aiscan,” built with DeepSeek and Kimi models, to chain the PHP vulnerability PHP2Shell with the Linux kernel privilege-escalation flaw “Fragnesia” and gain root on 25 PAN-OS firewalls and management servers in the United States, Japan, Taiwan and South Korea. Palo Alto Networks confirmed active exploitation of exposed management interfaces the same day.

#### “aiscan” built on DeepSeek and Kimi, and 19,000 fingerprinted firewalls

In Dream's analysis the operation began on 8 October when the attackers built the AI-enabled platform “aiscan”; by 23 October they had fingerprinted about 19,000 firewalls and narrowed the list to 240 targets. After intrusion they installed webshells, tunnelled internally, extracted credentials from LDAP and enumerated Active Directory. Three of the 25 compromised servers sat in a US government-affiliated environment, and Dream's Kfir Fleischer says at least two were US government servers. Attribution is bolstered by indicators such as naming conventions and the presence of authentication tied to a Chinese certificate authority, without confirmation from any public authority.

#### Chaining PHP2Shell with a kernel flaw

PHP2Shell is a PHP vulnerability with a CVSS severity of 9.2, and Fragnesia is a Linux kernel privilege-escalation flaw added to CISA's Known Exploited Vulnerabilities catalogue on 28 October. In a 29 October update Palo Alto Networks confirmed exploitation of internet-exposed PAN-OS management interfaces and shipped hotfixes for each release line, estimating that about 0.02% of the installed base is exposed.

Unit 42 meanwhile retracted its 23 October statement linking PHP2Shell to a European government agency. With potentially vulnerable PHP instances estimated above 22.6 million and only about 2.6% patched by one count, the same chain has ample room to spread to other appliances.

#### Victims including Japan, and the “AI-assisted” label

The inclusion of Japan, Taiwan and South Korea among the victims overlaps with the AI-enabled attack infrastructure targeting Japan reported in late October by TrendAI and CrowdStrike. Dream distinguishes this operation from autonomous attacks, calling it AI-assisted, with models supporting reconnaissance and target selection. The use of Chinese open-weight models as attack tooling adds a concrete case to debates over export controls and the regulation of open-weight models. For Japanese organisations the immediate tasks are checking exposed management interfaces and confirming that kernel fixes have been applied.

#### What to watch next

Advisories from CISA and Japan's JPCERT/CC; further disclosure by Palo Alto Networks of the scope of compromise; identification and disclosure of victim organisations in Japan.

#### Sources

- [S1] [Palo Alto Networks firewalls compromised in chained attacks using PHP2Shell](https://www.cybersecuritydive.com/news/palo-alto-networks-firewalls-compromised-chained-attacks-php2shell/873323/) — Cybersecurity Dive; 2026-10-30; reporting

- [S2] [Palo Alto Confirms Active PHP2Shell Exploitation](https://cybersecuritynews.com/palo-alto-php2shell-exploitation/) — Cybersecurity News; 2026-10-30; reporting

- [S3] [Chained PHP2Shell exploit roots 25 servers](https://dreamgroup.com/blog/chained-php2shell-exploit-roots-25-servers) — Dream; 2026-10-29; primary

## Briefs

### Taiwan's Asia Tech case: founder and wife detained on National Security Act suspicions as the immigration agency refuses delivery of 30 drones

Taichung prosecutors said the case began when Army acceptance inspection found parts stamped “CN,” leading to mid-October searches of the company and homes and the detention of founder Tu Ta-sen and his wife on National Security Act suspicions. Of the 32 units identified by the Interior Ministry, 30 are an undelivered National Immigration Agency order from Su Yuan Technology worth NT$2.75 million; the agency will refuse the company's and any Chinese-made products, and the ministry has not ruled out claiming damages. Defense Minister Wellington Koo has ordered an Army review and will announce results once confirmed.

- [S1] [亞拓科技無人機涉混用陸製晶片！內政部清查3機關有32架：不排除求償](https://www.ctwant.com/article/504647/) — CTWANT; 2026-10-31; reporting

- [S2] [顧部長：精進心輔作為 釐清無人機晶片來源](https://www.chinatimes.com/realtimenews/20261031000184-260417) — 中國時報; 2026-10-31; reporting

### Korean press reports Lazarus “Operation FlashHole” used Korean academic and media websites as C2, with post-quantum encryption and AI-generated résumés

Kaspersky's report of 21 October describes Lazarus approaching Korean firms through fake résumé sites, abusing the financial security software INISAFE CrossWeb EX for a multi-stage in-memory intrusion, and using ML-KEM-1024 post-quantum key exchange to protect its channels. On 1 November Korean outlets reported that the published indicators include a page of the Korean Society of Obstetrics and Gynecology used as a command server, which the society is now checking. The résumé photo is assessed as likely AI-generated, and at least three companies have been hit since July 2025.

- [S1] [Lazarus Operation FlashHole](https://securelist.com/lazarus-operation-flashhole/121541/) — Kaspersky Securelist; 2026-10-21; primary

- [S2] [북한 해킹조직, 국내 학회·언론사 웹사이트 악용 정황](https://www.sentv.co.kr/article/view/sentv202611010010) — 서울경제TV; 2026-11-01; reporting

- [S3] [가짜 이력서로 접근한 北 해킹조직 기업 홈피 장악해 '통신거점' 사용](https://www.kmib.co.kr/article/20261101501274) — 국민일보; 2026-11-01; reporting

### Russia's statistics agency hit by a multi-day cyberattack; Belarusian Cyber Partisans and Silent Crow claim 23 TB, Rosstat denies any leak

Rosstat acknowledged that its main and regional websites came under a coordinated, carefully planned attack from many locations between 29 October and 1 November, with the main site down on Saturday. The Belarusian Cyber Partisans and Silent Crow claim to have taken 23 TB including wartime losses, demographic and economic statistics and restricted documents, and defaced pages read “No one will escape accountability.” Rosstat says monitoring caught the attack early and that no databases or legally protected information leaked, so the theft claim remains self-reported for now.

- [S1] [Russia's Statistics Agency Hit by Cyberattack as Hackers Claim Massive Data Theft](https://www.themoscowtimes.com/2026/11/01/russias-statistics-agency-hit-by-cyberattack-as-hackers-claim-massive-data-theft-a94146) — The Moscow Times; 2026-11-01; reporting

### Cisco Talos reports an infostealer that has frontier models appraise stolen data to rank victims, switching models when refused

“VOLTA-AUTODOXER,” described by Talos on 31 October, harvests browser credentials, service sessions, SSH and cloud keys and wallets, then sends the bundle via OpenRouter to Anthropic's top model cast as an analyst drafting a target assessment, falling back to Moonshot's Kimi and then Google's Gemini when it detects refusal phrases. The output is saved as a document with sections on worth and operational weaknesses, and ranked dossiers appear to be sold via Telegram. The sample was first seen on 11 October, and no government or defence victims have been identified.

- [S1] [Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer](https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/) — Cisco Talos; 2026-10-31; primary

### US Navy presolicitation seeks “Super Intelligence” for SSN(X) arrangement design; expressions of interest due 6 November

NAVSEA's SSN(X) program office posted a presolicitation on 30 October for software to support the arrangement of propulsion plant, weapons, sensors, machinery and crew spaces, using the term “Super Intelligence” that the Navy mandated in October for AI. Non-proprietary expressions of interest are due 6 November, possibly followed by a formal call including an other-transaction authority. No funding or award count is stated; the first SSN(X) has slipped to FY2040 and CBO estimates about 8.7 billion dollars per boat.

- [S1] [US Navy seeks 'Super Intelligence' for SSN(X) submarine design](https://interestingengineering.com/military/us-navy-super-intelligence-ssnx-submarine-design) — Interesting Engineering; 2026-10-30; reporting

- [S2] [U.S. Navy seeks AI for SSN(X) submarine design as American shipyards struggle to build its fleet](https://www.armyrecognition.com/news/navy-news/2026/u-s-navy-seeks-ai-for-ssn-x-submarine-design-as-american-shipyards-struggle-to-build-its-fleet) — Army Recognition; 2026-10-30; reporting

### FCC releases the final text of its 800 MHz drone order with a footnote on SpaceX's adjacent-band application; effective 30 days after Federal Register publication

A day after its 29 October vote, the FCC released the text of Report and Order FCC 26-72 lifting the 1991 airborne ban in 824–849 and 869–894 MHz for drones. A new footnote notes SpaceX's 9 October application to acquire Grain and T-Mobile licences at 817–824 and 862–869 MHz and expects drones to coexist with potential mobile and direct-to-device operations. The 19 November meeting will vote on 450 MHz drone links and 24.45–24.65 GHz counter-drone radar.

- [S1] [FCC 800 MHz drone order final text adds SpaceX Starlink footnote](https://dronexl.co/2026/11/01/fcc-800-mhz-drone-order-final-text-spacex-starlink-footnote/) — DroneXL; 2026-11-01; reporting

- [S2] [Report and Order FCC 26-72](https://docs.fcc.gov/public/attachments/FCC-26-72A1.pdf) — FCC; 2026-10-30; primary

### Pakistan's ISPR chief calls the army “technology-savvy,” citing AI, quantum and cyber while declining to say whether AI picks targets in Afghanistan

In a Reuters interview recorded on 30 October and broadcast on 1 November, Lt. Gen. Ahmed Sharif Chaudhry said the army is incorporating AI, cyber, drone operations and long-range fires; asked whether AI identifies targets in Afghanistan he declined specifics but said AI, quantum and cyber can and should be used. He claimed Pakistan shot down dozens of sophisticated Indian drones in May 2025 and that drones from Afghanistan come with Indian sponsorship, claims India denies.

- [S1] ['We're a tech-savvy army': DG ISPR voices confidence in Pakistan's capabilities](https://www.dawn.com/news/2041391/) — Dawn; 2026-11-01; reporting

- [S2] [Pakistan capable of intercepting any drones sent from Afghanistan, India: DG ISPR](https://tribune.com.pk/story/2638130/) — The Express Tribune; 2026-11-02; reporting

### Saudi Arabia intercepts two suspected Houthi drones over Riyadh and the south; Egypt's parliament authorises military support and France and the UK deploy forces

Coalition spokesman Turki al-Malki said on 31 October that air defences intercepted and destroyed a hostile drone over the Riyadh region and later a second launched toward the Southern Region, with no injuries and no stated launcher for the Riyadh drone. The Houthis declared a maritime blockade of Saudi Arabia after striking Riyadh airport; Egypt's parliament has authorised military support to the kingdom, France is sending troops to protect an oil terminal, the UK is deploying fighters, and the US embassy ordered non-emergency staff out.

- [S1] [Saudi Arabia says at least 2 suspected Houthi drones shot down over kingdom](https://www.timesofisrael.com/saudi-arabia-says-at-least-2-suspected-houthi-drones-shot-down-over-kingdom/) — The Times of Israel（AFP）; 2026-10-31; reporting

### Iran's IRGC claims to have captured a US MQM-170 target drone and destroyed an unmanned surface vessel; CENTCOM says all USVs are accounted for

On 30 October the IRGC Navy claimed four tankers exploded in the Strait of Hormuz, one from a mine, and that it had shot down and captured a US MQM-170 drone and destroyed an unmanned surface vessel. UKMTO confirmed one vessel was struck by an unknown projectile about 40 nautical miles south of Lavan Island, while CENTCOM denied the USV claim, saying all its unmanned surface vessels are accounted for. On 31 October Mojtaba Khamenei was reported to have ordered a response without geographic limits.

- [S1] [After talks in Tehran, Pakistan military 'cautiously optimistic' on US-Iran mediation](https://www.timesofisrael.com/after-talks-in-tehran-pakistan-military-cautiously-optimistic-on-us-iran-mediation/) — The Times of Israel; 2026-10-30; reporting

- [S2] [Khamenei said to order military response without 'geographic limits' if US, Israel attack](https://www.timesofisrael.com/khamenei-orders-military-response-without-geographic-limits-if-us-israel-attacks/) — The Times of Israel; 2026-10-31; reporting

### Sankei reports the planned defence-industry corporation will likely be an independent administrative agency co-founded with METI, starting at about 300 staff and growing to 600, with legislation next year

Sankei reported on 1 November that the semi-public corporation envisaged in the revision of the three security documents will most likely be an independent administrative agency co-founded with the Ministry of Economy, Trade and Industry, chosen for its strength in industrial policy. A defence industrial base strategy is to be drafted by year-end, with a bill in next year's ordinary Diet session and launch within the next fiscal year. Staffing would start at about 300 and grow to 600, covering wartime surge production, expanded transfers including co-production and maintenance with like-minded countries, and new ways of fighting with drones and AI.

- [S1] [防衛産業の新法人、経産省と共同設立の独法が有力　来年度中の発足目指す](https://jp.headtopics.com/news/3845034907299871003158-89426168) — 産経新聞（Head Topics転載）; 2026-11-01; reporting

### Club Tourism reports a possible leak of about 4.32 million member records and an NTT Data subcontractor's SaaS exposes bond-buyer data, as supply-chain breaches continue in travel and finance

Club Tourism said on 31 October that an intrusion detected early on 30 October may have exposed about 4.32 million member records including names, nationality, dates of birth, addresses, phone numbers, emails and emergency contacts, and took its website offline. The same day NTT Data said a digital-securities SaaS operated by a subcontractor was accessed without authorisation and suspended its bondholder benefits service; Rakuten Securities cites 362 records, Kintetsu Group Holdings 2,678 and JR West two, with the subcontractor unnamed.

- [S1] [クラブツーリズム、不正アクセスで会員約432万件の情報漏えいの可能性](https://www.traicy.com/posts/20261101386570/) — TRAICY; 2026-11-01; reporting

- [S2] [クラブツーリズム不正アクセスの解説](https://innovatopia.jp/cyber-security/cyber-security-news/119597/) — innovatopia; 2026-11-01; reporting

- [S3] [楽天証券・JR西日本・近鉄GHDの社債購入者情報、NTTデータ委託先への不正アクセスで漏えいの可能性](https://www.inatei.com/entry/rakuten-securities-data-leak-ntt-data) — 本家いなてい; 2026-11-01; reporting

### Zelensky says the strike on the Rostov solid-propellant plant also hit a Pantsir-S2 and an S-400 command vehicle, estimating about 70 million dollars in damage

On 1 November President Zelensky said that the 31 October Neptune and drone strike on the Kamensk chemical combine also hit a Pantsir-S2 air-defence system and a command-and-control vehicle of the S-400 Triumf, estimating damage at about 70 million dollars. The General Staff had reported a warehouse fire and damage to four production units, and Russian officials have not commented. The General Staff's tally the same day puts Russian personnel losses in October at 54,400, the highest monthly figure.

- [S1] [Ukraine war latest (1 November live blog)](https://kyivindependent.com/ukraine-war-latest-russian-oil-refinery-in-kstovo-ablaze-after-reported-overnight-drone-strike/) — Kyiv Independent; 2026-11-01; reporting

- [S2] [Головне за ніч та ранок неділі, 1 листопада](https://lb.ua/society/2026/11/01/777756_golovne_nich_ranok_nedili_1.html) — LB.ua; 2026-11-01; reporting

### Binance says about 3 billion dollars of the 6.3 billion stolen has been frozen and 6.34 billion reserved, with no evidence of a key leak and a post-mortem promised

On the theft of about 6.3 billion dollars in roughly twenty minutes on 29 October, Binance says it suspended withdrawals, froze about 3 billion dollars with partners and reserved more than 6.34 billion to cover losses. As of 1 November the exchange said there was no evidence that wallet private keys had leaked and promised a technical post-mortem. TRM Labs attributes the theft with high confidence to North Korea's TraderTraitor, while Binance itself has not publicly confirmed the attribution.

- [S1] [Cybersecurity Awareness Month Ends With A $6.3 Billion Halloween Scare](https://www.forbes.com/sites/emilsayegh/2026/10/31/cybersecurity-awareness-month-ends-with-a-63-billion-halloween-scare/) — Forbes; 2026-10-31; reporting

## Upcoming dates

2026-11-02 · GMO Internet Group press conference on the intrusion (afternoon, Tokyo)

2026-11-02 · Next round of the CCW Review Conference chairs' closed-door consultations with Geneva missions on the draft negotiating mandate

2026-11-02 · Palantir Q3 2026 earnings (after US market close), with commentary on the government segment and Maven

2026-11-03 · US midterm elections; OpenAI's election operations centre runs until counting ends

2026-11-06 · Deadline for expressions of interest under NAVSEA's SSN(X) “Super Intelligence” presolicitation

2026-11-06 · Deadline for GMO condolence-payment and compensation claims

2026-11-06 · UNGA First Committee session ends; the LAWS and military-AI resolutions go to the December plenary

2026-11-06 · Japan's emergency package of measures against cyber attacks (targeted for the first week of November; date indicative)

2026-11-09 · Deadline for responses to NCIA's RFI on the NATO Integrated Cyber Defence Centre technology pillar

2026-11-09 · US Congress returns for the lame-duck session; a House vote on impeaching Defense Secretary Hegseth is expected

2026-11-11 · Deadline for five frontier-AI companies to answer the House Democratic ranking members' letter

2026-11-12 · Anthropic's 2026 Usage Policy, which explicitly bars guidance and control software for drones and other autonomous vehicles, takes effect

2026-11-12 · GMO GlobalSign's final report to the Personal Information Protection Commission on the Trust Login misconfiguration

2026-11-12 · UVID Dronetech 2026 (Tel Aviv), including legal and ethical issues of lethal autonomous drones

2026-11-13 · Deadline (17:00 JST) for letters of intent under ATLA's RFI on UAV warhead fuzes

2026-11-13 · OpenAI's deadline to answer the letter from twenty House Democrats

2026-11-16 · CCW Seventh Review Conference (16–20 November, Geneva), considering the LAWS GGE final report and the draft negotiating mandate with a 2028 horizon

2026-11-17 · NATO Edge (Izmir, 17–19 November), including NCIA's industry workshop on the Integrated Cyber Defence Centre

2026-11-19 · FCC open meeting votes on 450 MHz drone data links and 24.45–24.65 GHz drone-detection radar

2026-12-07 · China International Aviation & Aerospace Exhibition (Zhuhai, 7–13 December), with a dedicated unmanned and counter-UAS zone and heterogeneous drone swarms announced

2026-12-08 · K-DEX 2026 (Goyang, 8–11 December), five themes including drones and counter-drone