Every layer of the Tesla “autonomy” story is really a story about human labor being kept out of view.

Every layer has built up a paper trail of humans: sworn testimony, labor board complaints, federal crash filings, jury verdicts, and regulatory findings. The trail starts in October 2016. Tesla published a video of a Model X driving itself, captioned “The person in the driver’s seat is only there for legal reasons.” Ashok Elluswamy, Tesla’s director of Autopilot software, testified in a July 2022 deposition that the video was staged. The car followed a predetermined route built with 3D mapping. Drivers intervened during test runs. A test car hit a fence during attempts to film driverless parking. Asked whether the video showed what a production car could do at the time, he answered “It does not.”

Musk lied. The deposition records that the video was made at his request. He then promoted it as proof the car drove itself with zero human input, while his own engineers were intervening on a mapped route.

It was the opposite. Humans were required. The labor was industrial in scale. Autopilot depended on thousands of annotators labeling driving footage at Tesla’s Buffalo plant. Software tracked their keystrokes, their time per clip, and their time away from the desk. The annotators announced a union campaign on February 14, 2023. Tesla fired dozens of them the next day. The union charge to the National Labor Relations Board counted more than 30 dismissals. Tesla counted 27 and cited performance reviews. In May 2024 the NLRB’s Buffalo regional director filed a complaint alleging Tesla maintained a workplace technology policy designed to discourage organizing.

The plant itself is a public asset. New York taxpayers spent roughly $1 billion building and equipping the Buffalo factory, which the state rents to Tesla for a dollar a year. When solar production faltered and Panasonic pulled out in 2020, Tesla filled the building with Autopilot data labelers, whose headcount satisfies the jobs quota that holds off a $41 million penalty. The concealed labor is also the subsidy compliance.

By 2024 the concealed labor had moved deep into the product. The Optimus robots serving drinks at the October 10, 2024 We, Robot event were operated remotely by Tesla employees. Morgan Stanley confirmed the teleoperation in an analyst note the next day. The robotaxi service launched in Austin in June 2025 with a safety monitor in every car and a remote operations team behind them. NHTSA contacted Tesla within days after videos showed wrong-lane driving and erratic braking. Crash filings unredacted in May 2026 show two of the seventeen reported robotaxi collisions happened while a teleoperator was driving: over a curb into a metal fence in July 2025, into a construction barricade in January 2026. Tesla told lawmakers its remote operators drive only below 10 miles per hour. The crashes came at 8 and 9. Both followed the same sequence: the automated system stalled, a human took remote control, and the car hit a fixed object.

The courts are now attaching prices to the record.

Tesla settled the Huang case on the eve of trial in April 2024, amount sealed. On August 1, 2025 a Miami federal jury in Benavides v. Tesla awarded $129 million in compensatory damages and $200 million in punitive damages, finding Tesla 33 percent liable for the 2019 Key Largo crash that killed Naibel Benavides Leon. Tesla’s share came to roughly $243 million, the first federal jury verdict over a death with Autopilot engaged. On February 20, 2026 Judge Beth Bloom denied Tesla’s post-trial motions. In December 2025 the California DMV adopted an administrative law judge’s finding that the terms Autopilot and Full Self-Driving Capability are misleading and violate state law (Case Nos. 21-02188 and 21-02189). Just say fraud. Tesla revised its California marketing to avoid a 30 day license suspension, then sued the DMV to reverse the finding it had just complied with. A federal criminal investigation into the self-driving claims has been open since 2021.

The regulator’s arithmetic covers the same decade. NHTSA closed its Autopilot defect investigation in April 2024 after reviewing 956 crashes and finding a critical safety gap between what drivers expected of the system and what it could do. The reviewed crashes included 29 deaths. The December 2023 recall that concluded the probe covered 2,031,220 vehicles. Four months after the recall, NHTSA opened a new investigation into whether the remedy worked, citing 20 crashes in updated cars. Independent tracking counts at least 65 deaths where Autopilot or FSD was cited as a factor, and the special crash investigations continue: another opened in June 2026 after a Model 3 running Tesla’s supervised self-driving software hit a house in Katy, Texas, killing 76-year-old Martha Avila.

History repeats?

In 1770 Wolfgang von Kempelen unveiled the Mechanical Turk, a chess machine with a human master hidden in the cabinet. The machine drew the crowds. A man still did the work. Tesla runs the same exhibition at scale, with annotators in Buffalo and monitors and remote operators in Austin, all concealed from people drinking the Musk “technology” kool-aid. “Geturked” in German means fake.

The court record says Tesla sold fiction as capability for a decade, concealed the people who made the fiction run, and fired the ones who objected. The criminal file remains open because the sales deaths continue.

OpenAI published an open letter on Thursday with a little over 100 signatures. That’s not a lot. And when you run the numbers, it reveals some interesting patterns.

For example, there are four targeted audiences for a “what needs to happen next” section:

1. every organization
2. cybersecurity companies
3. governments
4. frontier AI companies

And then it differentiates who it considers victims:

1. hospitals
2. water treatment plants
3. infrastructure that powers the internet

Already it’s getting weird. Why aren’t governments in the victim list? Or what about every organization being in the list? I don’t get it and neither should you. This is not a letter that makes much sense.

Let me explain further. Two of the four targeted audiences signed: the vendors, nearly to a company, and a scattering of enterprise buyers who fall under “every organization.” Zero governments. Zero from any of the three named victim sectors.

That is to say, plainly, there is no hospital, no water utility, no government, no regulator, no standards body beyond the one that runs a federal ISAC, and no operator of the internet backbone the letter says is at risk, signing it.

The people asked to act did not sign. The letter was signed by the people selling the thing it says is needed, plus a few of their customers.

That’s the whole game. It’s perhaps the dumbest possible version OpenAI could have published. Either the victims were never asked, in which case this is a vendor letter to buyers dressed as a coalition, or they were asked and declined, in which case the people the forecast is all about do not believe in it. Either way, this letter doesn’t pass a basic sniff test.

Tables are great

I thought maybe it would be fun to list each sector by the distance between what the letter claims and who actually signed. Think of it as a buyers guide, so you can easily migrate away from the brands who signed this strange letter. Rank 1 is the largest gap.

Rank	Sector	What the letter says	Signed	Absent
1	Water, energy, utilities	Named victim, named funding recipient, named access recipient	none	American Water, Veolia, Suez, Xylem, Duke, Exelon, NextEra, E.ON, EDF, Enel, National Grid, Colonial Pipeline
2	Healthcare	Named victim, named access recipient	none	UnitedHealth / Change Healthcare, Ascension, HCA, Kaiser, Mayo, Epic, Oracle Health, Philips, GE HealthCare, Siemens Healthineers, McKesson, CVS, NHS
3	Government, regulators	Entire section 03: fund, coordinate, expand trusted access, impose costs	none	CISA, NSA, NCSC, BSI, ENISA, CAISI, UK AISI, NIST, FBI, Europol
4	OT / ICS security	The vendors who defend rows 1 and 2; “critical infrastructure supply chain manufacturers” named	none	Dragos, Claroty, Nozomi, Armis, Forescout, Schneider, Siemens, Rockwell, Honeywell, ABB, Emerson
5	AI security (securing the model, the agent, its data)	“Build observability and security tools, ensure agentic identities are traceable and accountable”	none independently; Protect AI, Lakera, Prompt Security, Robust Intelligence and CalypsoAI only via acquirers Palo Alto, Check Point, SentinelOne, Cisco and F5	Knostic, Wirken, Zenity, Noma, Pillar, Lasso, HiddenLayer, Mindgard, Aim Security, Cloud Security Alliance
6	Standards, coordination	“Verified fixes,” “private disclosure,” CVE throughput	Center for Internet Security	MITRE, CVE Program, FIRST, OWASP, OpenSSF, ISACA, ICANN, IETF, Internet Society, EFF
7	Telecom	Named victim: “infrastructure that powers the internet”	Deutsche Telekom, Lumen	AT&T, Verizon (Salt Typhoon victims), T-Mobile US, Orange, BT, Vodafone, Telefonica, NTT, Comcast, Nokia, Ericsson
8	Operating systems, open source	“Open-source maintainers” named as recipients; “longstanding bugs” is their code	Red Hat, Hugging Face	Linux Foundation (Glasswing partner), Canonical, SUSE, Apache, Debian, Mozilla, Python Software Foundation, OpenSSL, curl, Rust Foundation
9	Internet-scale observation	The only parties who could measure “far more widespread”	none	Shodan, Censys, GreyNoise, watchTowr, Shadowserver, abuse.ch
10	Logistics	Among the largest documented losses from one automated attack	Flexport	Maersk, FedEx, UPS, DHL, C.H. Robinson
11	Automotive	Sector-scale supply chain and OT exposure	General Motors	Ford, Stellantis, Toyota, Volkswagen, BMW, Mercedes, Tesla, Hyundai, Rivian
12	Insurance	Hold the loss data the letter never cites	Zurich, Marsh	Chubb, AIG, Beazley, Coalition, At-Bay, Resilience, Munich Re, Swiss Re, Lloyd’s, Aon, WTW, Gallagher
13	Identity, PKI	“Agentic identities traceable and accountable”	Okta, 1Password	DigiCert, Sectigo, Entrust, Let’s Encrypt / ISRG, GlobalSign, Ping, Yubico, Bitwarden
14	CDN, edge	“Infrastructure that powers the internet”	Cloudflare, Akamai, F5	Fastly, Imperva, Netlify, Bunny
15	Dev platforms, AI code	“Raise the security bar for AI-generated code”	Cognition, Lovable, Replit, Vercel, Factory, Figma	GitHub, GitLab, Cursor, JetBrains, Atlassian, StackBlitz, Sourcegraph
16	Consulting, integrators	“Hands-on support”	Accenture, Capgemini, Cognizant, KPMG, PwC, Oliver Wyman, IBM, Unisys, WWT	Deloitte, EY, McKinsey, BCG, Booz Allen, Leidos, SAIC, CACI, Kroll, Optiv, GuidePoint, TCS, Infosys, Wipro, Atos
17	Vuln research, bounty, audit	“Authorized testing, private disclosure”	HackerOne, Trail of Bits, SpecterOps, Cantina, Zero Day Initiative (via TrendAI)	Bugcrowd, Synack, Intigriti, NCC Group, Bishop Fox, IOActive, Cobalt, Semgrep (OpenAI grantee), iVerify (Trusted Access participant)
18	Banks	Buyers; Trusted Access participants	Capital One, Citi, Fifth Third, U.S. Bank, NAB, Nationwide	JPMorgan (Glasswing, Trusted Access), Bank of America, Goldman, Morgan Stanley, Wells Fargo, BNY, HSBC, Barclays, Deutsche Bank, Commerzbank, UBS, Santander, BNP, ING
19	Payments, market infrastructure	Buyers	Mastercard, Visa, FIS, Fiserv, DTCC, The Clearing House, Block, Robinhood	Amex, PayPal, Stripe, SWIFT, Nasdaq, ICE, CME, Deutsche Borse, Euroclear, Coinbase
20	Enterprise software	Buyers and patch sources	SAP, ServiceNow, Adobe, Snowflake, Elastic, Incident.io	Salesforce, Workday, Databricks, Intuit, Autodesk, Zoom, Dropbox, Box
21	Retail, commerce	Buyers	Shopify, GoDaddy	Amazon retail, Walmart, eBay, Target, Home Depot, Alibaba, Automattic
22	Credit bureaus	Buyers	TransUnion	Equifax, Experian
23	Funds, private equity	Capital behind the sellers	Citadel, ExodusPoint, Advent	BlackRock (Trusted Access), Thoma Bravo, Vista, KKR, Blackstone, Insight, a16z, Sequoia, Altimeter, Greylock, Lux, Battery, ICONIQ
24	Silicon	Compute	AMD, Arm, Micron, Broadcom	NVIDIA (Glasswing, Trusted Access), Intel, Qualcomm, TSMC, Samsung
25	Hyperscale cloud	Compute	AWS, Google, Microsoft, Oracle, IBM	Alibaba Cloud, Hetzner, OVH, Scaleway, DigitalOcean
26	Frontier labs	Authors; section 04 is their own commitments	OpenAI, Anthropic, Google, Microsoft	Meta, xAI, Mistral, Cohere, NVIDIA, Apple
27	Security incumbents	Sellers	CrowdStrike, Palo Alto, Zscaler, SentinelOne, Fortinet, Check Point, Cato, Sophos, Proofpoint, Tenable, Okta, Darktrace, Cisco, TrendAI (Trend Micro)	Rapid7, Qualys, Netskope, Wiz (Google), Arctic Wolf, Huntress, Tanium
28	AI-native offense and SOC startups	Sellers	XBOW, RunSybil, Tenzai, depthfirst, Calif, Cogent, Corridor, Octane, Prophet, Dropzone, Cotool, Outtake, Abnormal, Aikido, APIsec	Horizon3, Pentera, Sublime, Torq

I guess I could have expanded the list of people who didn’t sign. It’s massive. The letter really does speak to something strange. Rows 1 through 5 are the letter’s stated purpose with zero independent signatures. Zero. Row 26 is its authors and rows 27 and 28 are their sales channel, where suddenly there’s near-complete coverage.

The signature density is the total inverse of the stated priority.

Thoma Bravo owns three of the security signatories, Darktrace, Sophos and Proofpoint, but for some reason it did not sign. Maybe it’s waiting to see how bad the reaction is. I found it especially odd that Merck, FedEx and Maersk, the three companies with the largest documented losses from NotPetya, did not sign. They felt the pain and said no to this letter. AT&T and Verizon, the carriers Salt Typhoon lived inside for a year, did not sign. Lumen, named alongside them in the original reporting, did sign, which makes it a signatory with a documented state intrusion in its own network and nothing to say about it in the letter. MITRE, which runs CVE, and FIRST, which runs CVSS, are absent completely in a letter about surging vulnerability throughput. Because why?

When “global” means America

I went with 108 of the 116 signatories because eight could not be verified. I almost did this by city, because it’s basically all America, but let’s start with nations.

Country	Count	Signatories
United States	87	Abnormal AI, Accenture, Adobe, Advent International, Akamai, AMD, Anthropic, APIsec, AWS, Block, Broadcom, Calif, Cantina Security, Cape, Capital One, Center for Internet Security, Cisco, Citadel, Citi, Cloaked, Cloudflare, Cogent Security, Cognition, Cognizant, Corridor, Cotool, CrowdStrike, Dell, depthfirst, Dropzone.ai, DTCC, Equinix, EXA.ai, ExodusPoint, F5, Factory, Fifth Third Bank, Figma, FIS, Fiserv, Flexport, Fortinet, General Motors, GoDaddy, Google, HackerOne, Hugging Face, IBM, KPMG, Lumen Technologies, Marsh, Mastercard, Mercor, Micron, Microsoft, Obsidian Security, Octane Security, Okta, Oliver Wyman, OpenAI, Oracle, Outtake AI, Palo Alto Networks, Perplexity, Proofpoint, Prophet Security, Red Hat, Replit, Robinhood, RunSybil, SentinelOne, ServiceNow, Snowflake, Snyk, Socket, SpecterOps, Tenable, The Clearing House, Trail of Bits, TransUnion, U.S. Bank, Unisys, Vercel, Visa, WWT, XBOW, Zscaler
United Kingdom	6	Arm (SoftBank-owned), Darktrace (Thoma Bravo-owned), Incident.io, Nationwide Building Society, PwC, Sophos (Thoma Bravo-owned)
Israel	4	Cato Networks, Check Point, Cyera, Tenzai
Canada	2	1Password, Shopify
Germany	2	Deutsche Telekom, SAP
Netherlands	1	Elastic
Australia	1	National Australia Bank
Belgium	1	Aikido
France	1	Capgemini
Japan	1	TrendAI (Trend Micro)
Sweden	1	Lovable
Switzerland	1	Zurich Insurance Company

Eighty percent of verified signatories are American. That’s funny for a “global response” letter.

Israel supplies four, all security vendors, because of course. The European Union, where NIS2 and DORA already impose the incident reporting and remediation duties the letter asks for, supplies a measly six: two German, one Dutch, one Belgian, one French, one Swedish.

And then China, India, South Korea, Taiwan, Brazil and every African and Latin American state supply zero.

The letter calls for a “global response” and “new partnerships” and signs itself with 87 American companies, four Israeli security vendors, one Japanese antivirus company under a five-month-old name, and six companies from the European Union.

I guess it’s like how baseball has a world series, or so I’m told.

The cartel thing again

Section 03 asks governments to “expedite the expansion of trusted access programs.”

Section 04 commits frontier companies to “responsible model access.”

Ok, so those programs exist and the letter’s authors are the ones who run them. OpenAI’s Trusted Access for Cyber named its participants on April 16: Bank of America, BlackRock, BNY, Citi, Cisco, Cloudflare, CrowdStrike, Goldman Sachs, iVerify, JPMorgan Chase, Morgan Stanley, NVIDIA, Oracle, Palo Alto Networks, SpecterOps, US Bank and Zscaler, plus grant recipients Socket, Semgrep, Calif and Trail of Bits. Anthropic’s Project Glasswing named Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks on April 7. On August 10 OpenAI split Daybreak into Blue and Red tiers, both gating “limited-access frontier cyber models” to approved customers.

Overlay what’s being recommended in the letter with who is in and who is out.

Of 17 Trusted Access participants, nine signed and eight did not. Of the four grant recipients, Socket, Calif and Trail of Bits signed and Semgrep did not. Eight Glasswing partners signed and four did not. Apple, NVIDIA, JPMorgan and the Linux Foundation hold Mythos access and declined to endorse the letter. Bank of America, Goldman, Morgan Stanley, BlackRock, BNY and iVerify hold GPT-5.4-Cyber access and declined. The companies already inside are split on whether to ask governments to do this or not.

Only the companies selling are who signed unanimously.

The anchor for “in the coming months” is also suspect. Says who? OpenAI? Their Defender’s Window post describes an agentic collective that autonomously penetrated OpenAI research infrastructure and Hugging Face production infrastructure during an evaluation. Hugging Face signed. The letter cites the incident nowhere, and skips the detail that the agents got in by chaining unknown bugs with credentials already leaked on the internet. Half of that is a zero-day story. The other half is a password story.

The letter contains no incident data, no actual science or references to forecasts. Its signatories perhaps should have contributed theirs. F5 disclosed in October 2025 that a nation-state actor held persistent access to its BIG-IP development environment and took source code. Capital One paid an $80 million OCC penalty for the 2019 breach of 106 million records. TransUnion disclosed a 4.4 million-person breach in August 2025. Snowflake’s 2024 customer breaches, Okta’s 2023 support-system breach, Oracle’s disputed 2025 cloud breach, CrowdStrike’s July 2024 outage. Every one of those companies signed. None of their data appears in a letter that claims it can predict what the next few months hold.

Follow the money

Tenzai raised $75 million in seed funding in November 2025 on the pitch that enterprises spend five dollars on services for every dollar on product and AI can take the services market. Outtake’s $40 million Series B in January lists Satya Nadella, Nikesh Arora, Shyam Sankar and Trae Stephens as investors; OpenAI is a customer. The Microsoft CEO and the Palo Alto CEO hold personal equity in one signatory that sells to another. Trend Micro renamed its enterprise business TrendAI on March 23 and signed under the new name five months later. Center for Internet Security runs MS-ISAC, whose federal cooperative agreement CISA ended on September 30, 2025; the letter’s “fund cyber defense, starting with essential services” is its own budget request, and it is the only nonprofit I could verify on the list.

What the letter should have said

Section 04 commits frontier companies to “responsible model access.”

Section 03 asks governments to expedite “trusted access programs.”

Both route defensive capability through the labs themselves. Trusted Access for Cyber and Daybreak Blue and Red are a thing OpenAI is selling, and the letter asks governments to expedite it, two clauses after asking them to fund cyber defense.

The alternative architecture already exists and is totally absent from the list. An operator-run gateway puts the model choice, the credentials, the egress policy and the logs on the defender’s side of the wire. Smart, right?

Wirken, released as open source in February 2026, connects the same agent to Ollama on a local box or to Anthropic, OpenAI, Gemini, Bedrock or NIM through one policy layer the operator controls. The “observability and security tools” and “traceable, accountable agentic identities” the letter says you have to get from the frontier companies are elsewhere. Try using gateway functions shipped under an open-source license. A water utility can run them today without a license fee, let alone applying to a lab for trust.

That is the difference between the last two table rows versus the five at the top. The signatories are selling access to defense, by saying it comes from their model. The absent sectors need control that is durable and cost-effective, less dependency on vendors and more independence from them. The letter is an example of how not to write a letter.

What could go wrong? Trump has ordered lakes in America to be renamed Lake America. Every lake shall forthwith be known for what it is, to reduce confusion about where it is. This genius move is believed to be the kind of bold and decisive action to make travel far easier, while it rapidly increases American test scores in geography. Likewise, every state will be more appropriately named to United State, with capitals renamed to State Capital.