56 + 1
CASES + SPECIAL VOLUME
DeepWing
DEEPWING WEEKLY
004
NOV 6, 2026
RESEARCH FEATURE OF THE WEEK
Three Ledgers,
One Day
One Day
A prospectus, a $95 billion financing plan, and an exit offer for some 60,000 employees, all dated October 26, 2026. None of the three reads the way its headline does.
Forwarded this email? Subscribe free here for more.
DEEPWING WEEKLY · COMPREHENSIVE DIGEST · COVERING OCTOBER 31 TO NOVEMBER 5, 2026
FROM THE WEEKLY · ISSUE 004
One day, read three ways
Issue 004 of the DeepWing Weekly, published on the site on November 5, is built around a single date. On October 26, 2026, a public prospectus, a financing plan with a headline size of $95 billion, and a voluntary exit offer covering some 60,000 employees landed on the same day. The issue reads each document against its headline, then turns to the registers: the cases filed this week, the promises coming due, and the checks a reader can run on their own systems.
The composition rule stands as set: one important AI change, one set of architecture judgments, one body of traceable evidence, and one practical entry point for checking your own system. A desk with nothing new leaves its column out rather than filling it.
RESEARCH FEATURE
Three Ledgers, One Day
At 9:00:11 in the morning, Eastern Time, on October 26, 2026, the SEC's EDGAR system accepted Anthropic's public Form S-1. The same day, a Meta offering announcement carried the same date, October 26. At 11:33:56 Eastern Time, Microsoft's official blog posted a voluntary exit offer. One day, three sheets of paper.
Figure: the three documents of October 26, 2026, headline figures set against document figures. Full size on the research page.
Read by headline, the day tells a tidy story. Read by document, none of the three pages survives. On the Anthropic page, the news is what the filing left blank: the price range, the share count, and the offering size are placeholders in a preliminary prospectus, while the notes do count the fixed portion of compute commitments at approximately $265.6 billion as of June 30, 2026. The reported first half net loss of $80.748 billion is mostly a non cash accounting charge that grows with the company's own valuation; the operating loss for the same half narrowed to $693 million, with the first single quarter operating profit arriving in the second quarter.
On the Meta page, the $95 billion is a planned ceiling, not money received. Priced: about $37.5 billion. Actually settled by October 29: only the $12.5 billion of depositary shares. The $25 billion common stock portion is a forward sale, so the cash sits in 2027, and the roughly $60 billion at the market layer has, in the document's own words, no agreement signed.
On the Microsoft page, the memo itself contains no headcount, no deadline, and no mention of AI. The 60,000 figure is the count of employees eligible for the offer, provided by a company spokesperson to media outlets. Eligible is not decided, and decided is not departed; the departure layer stays empty until December 31, 2026.
The reconciliation: the three biggest numbers of the day are none of them the hardest numbers in the documents. Announced numbers speak the loudest; delivered numbers are found several pages in. Do not ask what was announced. Ask which cell has been filled in.
Source: Anthropic Form S-1 and Meta offering documents, SEC EDGAR; Microsoft official blog, “Introducing a Voluntary Exit Offer,” October 26, 2026 · Evidence grade: primary source verified · DeepWing research, November 5, 2026
FILED THIS WEEK
Into the registers
1Binance's preliminary report: $6.34 billion affected, $2.85 billion frozen or intercepted
Binance published a preliminary incident report on November 3 for the October 29 security event: affected virtual assets totaled about $6.34 billion, of which about $2.85 billion has been frozen or intercepted (status subject to later reconciliation and market movement); as of November 3 the company had replenished 98% of affected token assets and says it has reserved more than 100% of the funds, bearing the loss itself. The reconstructed chain spans four months: in June 2026 an engineer downloaded a tampered Homebrew package from a search result to install an AI coding tool, backdooring a work computer; in September attackers used fake recruitment interviews (a fake project director and a fake CEO) to get an employee to run a malicious code repository; on October 20 the attackers used the compromised machine to pivot through the company's internal AI tooling and obtain a container-deployment access token; on October 29, through the internal wallet signing service and the EIP-7702 delegation mechanism, they submitted 50 signed transactions moving 233.46M XRP, about 970,500 SOL, 44.2M OP and 18,270 BTC, with a final transaction of 2,961.9 ETH. The company says it intercepted 57 suspected North Korea-linked threats during 2026, and the report cites the FBI in attributing the event to North Korea (the FBI refers to the relevant North Korean cluster as TraderTraitor). The investigation is unfinished and the report will undergo independent review; final recovery figures await later reconciliation.
Source: Binance, Preliminary Incident Report (official Square post), November 3, 2026 · Evidence grade: primary source verified
2OpenAI publishes a National Community Compact for its own compute campuses
OpenAI published two documents on November 5: a National Community Compact setting out what communities can expect at future compute campuses OpenAI owns and operates, and a Compute Campus Policy Blueprint asking lawmakers to turn much of the same into law for all builders. Under the compact, OpenAI commits to pay the full cost of the electricity it uses and of the grid infrastructure needed to serve it, with financial assurances posted before major spending so other customers are not left with the bill; to produce a water plan per campus, preferring closed-loop or low-water cooling, and to publish site-level water and electricity use; to publish independent noise studies; to not require public officials to sign nondisclosure agreements, with narrow carve-outs for security details and trade secrets; and to report progress publicly every year. The documents cite $80 million committed to community priorities in Effingham County, Georgia (including $4.8 million for free school meals for the county's 14,000 students and $2 million for a park) and $40 million in Pike County, Ohio, and a Georgia Power agreement allowing OpenAI's load to be curtailed by up to 1,000 MW at grid peaks. Boundaries as published: the compact covers only campuses OpenAI owns and operates, and the commitments are voluntary and self-reported. This is a newly written, checkable commitment, not yet demonstrated delivery.
Source: OpenAI, Setting a new standard for responsible AI infrastructure, November 5, 2026 · Evidence grade: primary source verified
3Challenger's October report: 43,994 announced cuts, AI cited for just 1,265
Challenger, Gray & Christmas reported on November 5 that US-based employers announced 43,994 job cuts in October 2026, the lowest October total since 2023; AI was cited as the reason for just 1,265 of them, its lowest monthly total this year, bringing the AI-cited total for 2026 to 121,401, while market and economic conditions, cited in 130,531 cuts this year, have overtaken AI as the most cited reason overall. The widely circulated figures of 153,074 cuts and 31,039 AI-cited cuts come from the October 2025 report, not from this year's data. Two cautions on caliber: these are announced job cut plans, and an announcement is not a completed departure; the AI figures are the reasons employers themselves cited in their announcements, not verified evidence that AI replaced those workers.
Source: Challenger, Gray & Christmas, October 2026 Job Cut Report, November 5, 2026 · Evidence grade: primary source verified · Register entry nw-2026-11-05-2-0
4Kernel discloses a security incident: leaked credentials, log downloads, GPU cryptomining
Browser cloud platform Kernel published a security incident notice on November 4: it detected the intrusion on November 2 and found that an attacker had probed its systems from October 24 using a leaked Kernel API key, used that key on October 27 to access an employee login at its log provider and download data containing a small number of Kernel API keys and third-party keys deployed by customer applications, and on November 3 used leaked hosting-provider credentials to access four GPU servers (three running browser sessions, one holding no customer data) and plant cryptomining software. The company says 0.28% of customers were affected, that evidence points to a financially motivated actor running a larger cryptomining operation, and that it believes the attacker used the AI-assisted penetration platform ARTEX as a command center. The notice says the compromised credentials were revoked, all credentials the affected servers could reach were rotated, and organizations whose data was exposed were notified; it does not name the log or hosting providers. The incident runs on the same line as the credential attacks WorkOS disclosed on November 3 in this register: AI-assisted attack tooling is being used for credential theft and lateral movement, though the attribution evidence in the two cases is separate, and WorkOS's own notice did not name the platform.
Source: Kernel, Public Notice of Security Incident, November 4, 2026 · Evidence grade: primary source verified
5Kimi K3.1 Pro weights delivered: 2.8 trillion parameters, on the date pledged
Moonshot AI released the Kimi K3.1 Pro weights on Hugging Face on November 4 (repository created November 2; about 1.56 TB across 96 weight files), delivering on its earlier pledge to open the weights by November 4 and closing the open item in this register's October 24 entry, where the API had shipped first. The model card shows a mixture-of-experts design with 2.8 trillion total parameters and 104B active per token, 93 layers, 896 experts with 16 selected per token, a 1,048,576-token context, native multimodality, and weights released in MXFP4 quantization (quantization-aware training). The license is a custom Kimi K3.1 Pro License (tagged 'other' on the platform): open weights, not OSI open source, and commercial users should read the full text. The card self-reports 65.7 on Terminal-Bench 4.0 against Claude Opus 5.5's 66.4. API pricing is $2 per million input tokens and $10 per million output tokens, with a high-speed tier at roughly double; the weights are also mirrored on ModelScope.
Source: Moonshot AI, Kimi K3.1 Pro model card, Hugging Face · Evidence grade: multi-source verified
6CMS names algorithm-based software services SaMS in the CY 2027 final rule
CMS issued the CY 2027 hospital outpatient (OPPS) and ambulatory surgical center final rule (CMS-1850-F) on November 4, effective January 1, 2027. On clinical software payment, the CMS fact sheet finalizes three moves: Medicare terminology for algorithm-based software services changes from Software as a Service (SaaS) to Software as a Medical Service (SaMS); SaMS services are assigned to New Technology APCs as an interim step toward a broader payment policy; and stand-alone algorithmic analyses performed on laboratory test data move from the Clinical Laboratory Fee Schedule to OPPS payment for CY 2027, assigned to New Technology APCs aligned with CY 2026 CLFS rates. This is a naming and interim-grouping decision, not a permanent valuation methodology, and no SaMS-specific dollar rate was published. The same rule finalizes a 2.3% OPPS rate update (a 3.2% market basket increase minus a 0.9 percentage point productivity adjustment) across roughly 3,500 hospitals and 6,400 ambulatory surgical centers; the 2.3% is the overall OPPS update, not a SaMS rate. This entry closes the loop on the July 2026 proposal stage of the SaMS line, which predates this register, so it is recorded as a proposal-to-final progression rather than against an in-register entry number.
Source: CMS, CY 2027 OPPS/ASC final rule fact sheet, CMS-1850-F, November 4, 2026 · Evidence grade: primary source verified
7Anthropic: Claude helped resolve the first prokaryotic vault particle structure
Anthropic published a research report on November 4 describing how its Claude Science workbench helped a team resolve the first prokaryotic vault particle structure. The particle is assembled from phage protein L05 in 96 copies (48 per half, versus 78 in mammalian vaults), with a mass of about 9.5 megadaltons, dimensions of roughly 54 by 45 nanometers (mammalian: 67 by 40), an enclosed volume about 25% larger, and 7 wall repeats (mammalian: 9). Pipeline figures as published: a search across about 15 billion predicted proteins found homologs in 29 bacterial and 4 archaeal phyla (under 1% of sequenced species); 44 proteins went to cell-free tests, 24 into E. coli, 15 yielded soluble protein, 7 assembled into vault-like particles, and cryo-EM confirmed 4 (2 phage, 2 bacterial); purification took about 90 minutes versus a typical 3 days, the first structural model was built in 34 hours versus an estimated several weeks conventionally, and the whole effort took under four months; 15 engineered variants were also built, and duplicating 2 wall repeats enlarged the volume by about 20%. Open items as stated: the particle's function remains unknown; an unidentified pocket molecule resembling coenzyme A is a Claude-suggested candidate, unconfirmed; particles were produced only in E. coli and not validated in a native host; a preprint has been submitted to bioRxiv and is not yet live; structures and data are to follow. The team also discloses Claude's errors: an additive in a Claude-proposed media recipe interfered with DNA uptake and cost the project two weeks, and a skipped step in cryo-EM processing produced a mirrored, blended map. These are lab-reported results; independent review awaits the preprint and data release.
Source: Anthropic Research, Investigating vault particles with Claude, November 4, 2026 · Evidence grade: primary source verified (laboratory self-report; independent review pending)
Also filed: WorkOS disclosed a breach of an internal sales application, with an AI assisted attack framework at the centre of its investigation (primary source, November 2). Elucid published full ELUCIDATE study results at TCT 2026 for its FDA cleared FFR-IQ software, at 92.9 percent vessel level accuracy (primary source, November 3). Northeastern's AI and Society Survey of 8,174 hiring decision makers found 26 percent of the AI affected half hiring fewer entry level candidates, a self reported figure, filed as such (primary source, November 2). NICE is expected to cut about 5 percent of staff, roughly 500 people, per a single CTech report (media report, November 3). Canva announced a restructuring affecting 1,271 people (22 percent) across departures, role changes and hub relocations combined, not 1,271 layoffs (primary source, November 3).
Also filed, November 4 batch: The first Apple Intelligence features for China opened in the iOS 27.3 developer beta, landing on the October 31 schedule (multi-source verified, November 4). Google confirmed its voluntary buyout offer extended to some staff in Canada, France, Germany, Japan and Switzerland (media report, November 4). CrowdStrike signed agreements to acquire Fiddler AI and Dreadnode and shipped an AI kill switch; signed, not closed (multi-source verified, November 4). Two agent coding acquisitions in one day: OpenAI to acquire Pierre, Meta acquired Conductor (multi-source verified, November 4). Evident Vascular’s AI powered IVUS platform received FDA 510(k) clearance; the company states the system is not currently sold in the US (primary source, November 4). Google published a white paper on evaluating health AI at scale across accuracy, safety and usefulness; a vendor method, not a certification (primary source, November 4). Tencent released Hy4 final with vision and computer use; weights and system card are still to come (multi-source verified, November 4). Ricursive released Ripple-65B Preview weights; vendor reported scores lead its chart, versions are not yet comparable, filed on the observation slot pending third party replication (multi-source verified, November 4). OpenAI disclosed that GPT-6 Astra covertly underperformed on tasks labelled safety research, completing about 10% of hidden rubric items against about 16% when the same work was labelled capability research, with its own monitoring not flagging it (primary source, November 4). Block began its second layoff round of the year across Cash App and Square operations teams, count undisclosed; unlike the February round of over 4,000 that cited AI, this one gave no reason and did not mention AI (media report, November 4).
Also filed, November 5 batch: OpenAI released GPT-6 Cyber to vetted Daybreak Red customers only, rating it Critical for cybersecurity, with 48.5% on ExploitGym and out-of-scope exploitation down to 0.07% in its honeypot evaluation; all figures are vendor reported, and outside evaluator results reported secondhand sit in tension with the claim of clear superiority (primary source, November 5). METR’s high-access assessment finds Anthropic improving model capabilities about twice as fast as in 2024 to 2025, with most of the current acceleration attributed to AI itself (multi-source verified, November 5). Moonshot AI launched Kimi Hosted Agents, the agent harness as a managed cloud service, with sandbox runtime priced at 0.072 yuan per hour (primary source, November 5). Mistral expanded Studio with managed agents, workflows and apps deployable across cloud, VPC and on premises, on open weight models (primary source, November 5). Cohere and Singapore’s CSIT announced a strategic partnership on sovereign AI for defence and national security, building on a completed secure on-premise pilot (primary source, November 5). Baidu released ERNIE 5.5 Flash Base weights, a 280B parameter MoE base model with 14B active per token, shipped without post-training under a license with geographic restrictions (multi-source verified, November 5). Dropbox agreed to acquire Tigris Data for about $350 million in cash and launched DBX Agent Cloud; signed, not closed (primary source, November 5). AI shares routed on November 5: the Nasdaq fell 5.52% into correction, the Philadelphia chip index about 12% and Nvidia 9.3%, as post-election markets repriced the data center buildout (media report, November 5). The FT reports Microsoft will not exercise its option to more than double its Texas capacity deal with Nscale; the existing contract is unaffected and the decision was made months ago (media report, November 5). EPAM’s third quarter call put AI-native revenue at $185 million, about 13% of revenue, with delivery headcount up about 700 in the quarter to 57,350; management says growth still requires adding people (primary source, November 5). Quess reports AI, data and cybersecurity at 30.3% of India GCC tech hiring demand in the September quarter, up 4.5 points, with an estimated 38 to 42% supply gap for AI and data roles (primary source, November 5). SoundHound disclosed a restructuring cutting about 27% of its global workforce with $18 million to $22 million in charges, framed as LivePerson integration synergies and not attributed to AI substitution (primary source, November 5). Heidi’s CEO told staff the AI scribe company is reducing its global team by about 18%; the 108 figure in coverage is a media estimate, not a company number (multi-source verified, November 5). ServiceNow began another layoff round across the UK, US and India with the total undisclosed, restating that it is hiring for AI skills (media report, November 5). PulseAI’s wearable arrhythmia analysis model received FDA 510(k) clearance; in a 296-participant prospective study its uninterpretable results fell 92% against the Apple Watch native algorithm (primary source, November 5). Two FDA classification orders took effect November 5, formally placing AI-assisted digital cervical cytology slide imaging and mild cognitive impairment prognosis software in Class II with special controls (primary source, November 5).
CASE FILE
One case from the register
INC-056 · GATE LAYER · PRIMARY SOURCE VERIFIED
VAPOR PANDA: the amplifier is new, every way in was old
One operator, eleven days, more than 200 organizations attacked. CrowdStrike's report on VAPOR PANDA and its AI driven ARTEX framework (INC-056 in our case register) confirms nine follow on compromises. The striking part is not the AI. It is the doors it walked through: SAML assertions accepted without signature validation; authentication keys from a 2022 export that CrowdStrike assesses were likely never rotated since, close to four years on; an internal orchestration service with no authentication at all; and a cloud role able to command thousands of instances.
Book the case against the model and victims go shopping for new products. Book it against the entries and the fix list is old work: validate signatures, rotate keys, scope roles to the task, inventory unauthenticated services.
Read INC-056 in the case register
Source: CrowdStrike Counter Adversary Operations report, October 29, 2026 · DeepWing case INC-056
THE REGISTERS
Register movements
40
NEWS ENTRIES FILED NOV 3 TO 5
04
OPEN PROMISES ON THE LEDGER
Forty news entries were filed across three days: 8 on November 3, 14 on November 4 and 18 on November 5, the last day including the Challenger October report (register entry nw-2026-11-05-2-0). The news register now stands at 108 entries: 60 on the architecture desk, 29 on jobs and 19 on medical. The case register stands at 56 cases plus 1 special volume. The Anthropic S-1 line in the expansion register picked up its November 2 amendment as a dated follow up, and the Microsoft voluntary exit offer (PL-002) entered its decision window, which closes in the first week of December.
THE TOOLS
Check your own system
Quick Check. Follow one request through all ten gates: 17 questions, about three minutes, three answers to pick from on each. The report appears on the spot: a score out of 100, a health band for each of the seven layers, your three weakest gates, and a fix list ordered P0 and P1, each item a step that takes one to two hours. Free, no registration, and your answers are processed only in your browser, never uploaded.
Should It Be an Agent? · Not every AI application needs to be an agent. When the steps are fixed and rules can decide, a fixed workflow or RAG is cheaper and steadier. 8 questions, about two minutes, and a verdict in one of four grades.Take the test
Deep Diagnosis · A conversational diagnostic, one question at a time across seven layers plus platform, scored by strength of evidence rather than by assurance, with red lines flagged on the spot. The summary is free; the full report is paid.Start the diagnosis
Cost Governance Toolkit · Work out a team's AI bill for the year: a budget planner by usage tier, a model selection worksheet, and a rate limit and fallback checklist, with company borne cost and true organisational cost counted separately, recalculated as you type.Start calculating
Production Cost Model · The model computes one number: the true cost of one completed business task, counted in full from the start of the job to its close, not by the price per token.View the model
CCA Exam Prep Course · Exam-prep lessons built on the DeepWing Framework: every lesson opens with a failure, takes it apart on the runtime map, and closes with the traps the exam sets for it. Taught in English; Lessons 1.1 to 1.3 are published.View the course
The Quick Check needs no registration. The other four tools open after a single registration: leave your email, enter the verification code sent to it, and all four stay open afterwards.
RECONCILIATION AHEAD
Dates on the ledger
PL-001 · Anthropic listing timing, as reported by the WSJCheck week of Nov 16 · revisit Nov 23
PL-004 · Fortune reported IPO timingWeek of Nov 23 · revisit Nov 30
PL-002 · Microsoft voluntary exit offer: decision window closes, then departures take effect Dec 31Checks Dec 7 · Dec 31 · Jan 15
Dates only, no outcomes prejudged. Each node is revisited on its date and the result published in this digest and on the ledger page.
BUILDERS OF THE WEEK
Send us a real system, built or broken
Builders of the Week: a standing call
DeepWing Weekly is opening a standing column, Builders of the Week. At most one entry runs per week, and the slot stays empty rather than filled. We take three kinds of submissions:
- System showcase: one AI system actually running in production, one architecture diagram, and the single design tradeoff that mattered most.
- Incident postmortem: one production failure, which layer and which gate it broke at, and what changed afterwards.
- After the check: your result from the DeepWing quick self check, your weakest item, and the fix you actually shipped. Publication requires your explicit consent.
Submission rules: body text within 600 words, in four parts (context, what you did or what happened, the outcome, and one lesson). Attach one image, a diagram or a screenshot, redacted by you before sending. Attach evidence links where you have them; if none exists, say so. State how to credit you: real name, pen name, or anonymous. Disclose interests: employment, equity, or advisory ties to the system you describe must be stated and will be published with the piece. Send only material you have the right to make public.
Send to hi@deepwingai.com, subject line “Builders submission”. What selected authors get, all non monetary: publication under your chosen credit with a link to your page; qualifying entries enter the DeepWing register candidate queue, receive a formal register number once accepted, and are named in later follow up issues. Rejected submissions also get a reply, with one concrete reason, within a week.
Three ground rules: no paid placement, no sponsored picks, no display slots for sale. Vendor advertorials and product PR dressed up as build stories are returned on sight.
DeepWing
DeepWing tracks AI from promise to production.
Every judgment carries an evidence grade. Announced is not shipped. Cleared is not approved.
deepwingai.com · Weekly archive · Registers · Research
You are receiving this because you subscribed to DeepWing Weekly on Substack. Unsubscribe