OpenAI’s summer of claims is turning into a fall with a citation problem.

The company called their Hugging Face breach unprecedented, called their Black Hat account of it a watershed, wrote that ten mathematics problems had seen no progress for a decade, let a “secret technique” story about Astra run for a day, and launched Astra claiming the first model to cross its own cyber threshold. And yet every single one of these is not new, because there is an easily proven prior record. In two cases the prior record is proven by OpenAI’s own documentation of it.

Precedented

OpenAI models in July under evaluation had their safety filters removed and were put in a sandbox with internet access disabled. The models exploited a zero-day in the package registry proxy, moved laterally until they reached a node with internet, and then broke into Hugging Face production to cheat the answers to the test. OpenAI’s description of this was “unprecedented.” I wrote the five whys the next day, explaining what really happened. The chain began with a file that was trusted without a reason to be trusted. Cliff Stoll published the genre in 1989. I gave the BSidesLV talk 23 years later, in 2012, on data as the attack vector for AI. The failure had at least fourteen years of public knowledge before OpenAI’s engineers said it was new in their own evaluation harness.

Watershed

OpenAI in August pumped their story at Black Hat. The presentation treated a sandbox design failure as fixed by rebuilding the compromised service in the same wrong way, and so the agents rebuilt their coordination channel within days. Incident response frameworks require root cause removal. OpenAI presented their inability to perform basic security operations, and basic incident response, as a wakeup in the field. The follow-up was no better.

A decade behind

OpenAI published on 1 August ten mathematics results attributed to an internal Astra model. The release claimed these are problems that “have been open and seen no progress on the main result for at least a decade.”

It’s not a claim that is hard to check. Stephen Miller of Yeshiva University found the sphere-packing proof resting on an argument from his own 2016 paper. Francesco Fournier-Facio of Cambridge found the non-sofic group construction assembled from Gábor Kun’s 2016 paper and the 2019 Kun-Thom paper.

OpenAI’s paper actually cites all three. Cohn-Miller 2016 appears once, for a preliminary reduction step, and Miller says the argument the proof hinges on came from that same paper and was presented as the model’s own. Kun and Kun-Thom are cited in the summary of the non-sofic result. It’s just that their PR describing the paper erased the prior work entirely. Scientific American ran a scathing indictment of OpenAI five days later. Miller called the pattern systematic and put it under research misconduct.

OpenAI edited their page to say instead each result “resolves or makes substantial progress on a long-standing open problem,” under the original date, with no note about being caught.

The same release carries a section on responsibility to the mathematical community. It argues that attribution should reflect how a result was produced, and that claiming human authorship for an AI-generated proof would misrepresent the work.

Sheesh.

Credit between the system and OpenAI’s own staff is settled in that paragraph. Credit to the 2016 and 2019 authors was handled by the sentence above it. An OpenAI spokesperson told Scientific American the company meets the standards expected of human mathematicians. A human mathematician who submits a proof on a 2016 argument credited only for a preliminary step, and describes the field as stalled for ten years, is facing misconduct charges.

Open secret

Four weeks later The Information reported a “secret technique” inside Astra. I traced it the same day: recurrent depth, Graves 2016, Dehghani 2018, Giannou 2023, Geiping 2025 with released weights, and Nanbeige shipping it under Apache-2.0 in July with forty thousand downloads a month. OpenAI, Anthropic and Google DeepMind had cited the Geiping paper by name in their July 2025 chain-of-thought monitorability statement as a risk to document. In that case OpenAI’s own chief scientist publicly rejected the novelty claim being made.

Opposite of first

On 3 September OpenAI launched GPT-6 Astra as the first model to meet the critical cybersecurity threshold of its own preparedness framework. Easy to say, of course, when the threshold is OpenAI’s to say about OpenAI. But everyone knows Anthropic gated Mythos for cyber capability in April. The launch post reports two zero-day vulnerabilities found during evaluation. The July incident report had already described the models chaining zero-days to leave the sandbox.

So vain, so lame

Five claims in ten weeks. Five flops.

One came from reporters working from an anonymous source, and the vendor’s own scientist loudly disowned it. Four came from OpenAI.

The July incident was a repeat of fourteen years of published defense literature. The August paper was 2016 and 2019 mathematics that the paper itself cites. The September architecture was 2016 machine learning that OpenAI itself cited back in 2025.

The company’s PR calls everything new, even when the company’s own documents record earlier work.

OpenAI states the mathematics manuscripts were prepared by humans working with the model. Miller’s 2016 preprint with Henry Cohn, arXiv 1603.04759, is dated March 2016 and sits in the paper’s own bibliography. Kun 2016 and Kun-Thom 2019 are on arXiv. Graves 2016 is on arXiv. Stoll’s book has been in print since 1989. All of it was available to the models, let alone the people who wrote “unprecedented” in July, “at least a decade” in August, and “first” in September.

The mathematics release said the work was roughly $2,000 in tokens. The architecture report framed the same model needing roughly $600 billion in annual capital expenditure. That’s a lot of money wasted, especially when you realize the prior work it plagiarized cost nothing to read.

Tesla told a Florida court it could find no Test Incident Reports from its underbody impact testing. Its own test protocols require a report, photos and video for every run. Well, that doesn’t make any sense, does it? Could Tesla be lying in court, again? Then its own witness, Adam White, testified the reports can be located by clicking hyperlinks in the company’s system.

Circuit Judge Michael A. Robinson called the claim “not credible”, probably because court hates calling people liars, and wrote that it “appears to have been a willful and/or intentional misrepresentation.” Liar, liar, why can’t we just call Tesla the liar it has been and continues to prove itself?

That was the first sanction in Alcala v. Tesla Florida. The second was issued August 19. Trial is set for November 30.

Underbody Underengineering

Nicholas Garcia, 20, drove a 2021 Model 3 through the intersection of Alhambra Circle and Coral Way in Coral Gables with Jazmin Alcala, 19, in the passenger seat. The underbody struck a crest in the road. The battery pack ignited. The car veered onto the median, glanced off one tree and struck a second with its passenger side. Both occupants died in the fire. The date was September 13, 2021.

Garcia had owned the car for about six days. Four days before the crash he had taken it to Tesla service for steering and suspension complaints. Six weeks after the crash, Tesla filed recall 21V-835 with NHTSA, covering 2,791 Model 3 and Model Y vehicles whose front suspension lateral link fasteners could loosen and separate from the subframe.

The NTSB preliminary report recovered five seconds of data from the event recorder. Maximum recorded speed was 90 mph on a 30 mph street, accelerator applied to beat a yellow light, service brake off. Tesla’s defense rests on that data. The family’s case rests on how the underbody and battery pack respond to road contact.

Tesla’s own tests address that question, and Tesla has spent three years fighting to prevent transparency into it.

Willful Disregard for Safety

The family filed in October 2021. Here’s the discovery record:

September 20, 2023. The court ordered Tesla to produce an engineer responsible for stability control. Tesla produced Daniel Wood, who testified that the engineer “personally responsible for the stability control” would be better placed to answer. Tesla designated no one else. Judge Robinson found this “a direct violation” of the order.

November 6, 2023. The court compelled testing documents for real-world driving situations, including speed bumps and uneven surfaces. Tesla’s counsel told the court everything responsive had already been produced. Over the following year Tesla produced no additional testing documents. The withheld material included the Sine Wave Test, which Judge Robinson found “substantially similar to the crest in the roadway that was involved in the subject incident.”

March 7, 2025. Five days before a scheduled deposition, Tesla produced 8,100 pages of test reports. The documents state on their face that the tests simulate conditions in which the battery pack and underbody make direct contact with the ground, bumps, curbs and other objects, to reproduce severe unintentional customer usage constituting an accident.

June 12, 2025. Tesla’s counsel wrote that Tesla “did not locate any TIRs.” The court found the opposite, that Tesla “was in fact in possession of thousands of pages of TIRs.”

Just call them liars, seriously. Nothing they say can be trusted. It’s like the courts don’t yet have the power to shut the whole thing down? How long did Enron stay in business?

July 2025. Four days before the sanctions hearing, Tesla produced 123,000 pages with metadata and file names stripped. Judge Robinson found this was done intentionally and that the acts “were intended to make the review and use of these materials more difficult, time consuming and expensive for the Plaintiffs.”

Imagine a company that knows it’s killing its own customers, and then tries to pull as much time and money as possible away from anyone who complains.

October 24, 2025. Judge Robinson granted the first motion for sanctions. He found Tesla “acted willfully or with contumacious and deliberate disregard” for two court orders, awarded the plaintiffs fees and costs including expert time spent reviewing the July production, and warned that continued violations could result in Tesla’s pleadings being struck.

So many warnings, so much time.

The Tracker

A separate dispute concerns Tesla’s internal incident tracker. Tesla claims the file was assembled at the direction of counsel and is privileged. The plaintiffs argue the TREAD Act requires Tesla to report similar incidents to federal regulators, which places the tracker outside privilege. A special magistrate ruled that Tesla’s witness must answer who has access to the tracker and what it contains. Judge Robinson affirmed over Tesla’s objections, with the contents themselves protected pending appeal. Tesla has said it will seek review from the Fourth District Court of Appeal.

Fraud: Second Sanction

The family filed a third motion for sanctions in May 2026. The motion and Tesla’s response each exceed 300 pages. Both are redacted in full as confidential. You would think Tesla is building some kind of military secret, like AI-based drones, given how they treat their products as super secret squirrel stuff. It reminds me of the kids in “Teenage Wasteland” who jumped the “no trespass” signs at the Wallkill landfill and exposed a mafia toxic dumping operation in upstate New York, with Ford Motor Company among the dumpers.

Retired Circuit Judge Charles M. Greene, sitting as special magistrate, heard the motion August 14 for three and a half hours, most of it closed to the public. Todd Falzone of Kelley Uustal told the court Tesla had violated three orders, paid monetary sanctions and continued to withhold. The production to date covers component testing. Vehicle-level testing under roadway conditions has yet to be produced. Franklin Sato, for the Garcia estate, called Tesla’s conduct “tantamount to fraud.”

Val Leppert of King & Spalding answered for Tesla. Eleven thousand documents produced, 9,000 of them testing records. Engineers made available for deposition. One hundred fifty hours spent building a mapping tool linking documents. Their production failures were “copying errors, misunderstandings and false starts.”

“We have not been perfect. There has not been any intentional withholding.”

The company that builds a car on the claims that it can automate better than humans, that it can define safety as a baseline, flails and falls down on basic document delivery tasks and begs for forgiveness. Yeah, nobody should trust that company with anything related to life or death decisions privately let alone on public roads.

Judge Robinson’s October order found the withholding willful and intentional and described a misrepresentation to the court.

Judge Greene issued his recommendation five days after the hearing. Tesla’s pleadings stand. He recommended an adverse jury instruction at trial, a further award of attorney fees, and compliance with the outstanding discovery. According to the family’s counsel, the instruction tells the jury Tesla knew its own testing had shown failures similar to the one at Coral Way.

That sounds right to me. Tesla has known since 2013 that AI was a wildly unstable theory, a pipe dream, and by 2016 knew Musk was lying about it to the press, even as two fatalities proved the system too dangerous for production. In March 2018 Uber killed a pedestrian in America and suspended all road testing the same day. In April 2018 Tesla killed a pedestrian in Japan, and two months later raised the price of Full Self-Driving from $3,000 to $5,000 while expanding the program. Most Americans haven’t heard about it because Tesla litigated to keep it all in Japanese. The constant rise in crashes and rise in fatalities came after they knew it would.

Going Back to Grimshaw

The Ford Pinto verdict rested on discovery. Ford’s rear-impact crash tests showed the fuel tank rupturing at moderate speeds. Ford produced the tests and a jury saw them. In Grimshaw v. Ford Motor Co., 119 Cal.App.3d 757 (1981), the Court of Appeal upheld punitive damages on the basis that the tests established what management knew.

Tesla’s approach in Alcala differs at the production step. Four years of orders, a production timed to a deposition, a second production timed to a hearing with filenames removed, a letter denying the existence of reports its own witness said were one click away, and a sealed 300-page attempt to call all of that purely accidental. It’s like they don’t believe in the laws of physics or the laws of America, operating above the laws and unaccountable while those who trust them are burned up or silenced.

Tesla’s litigation position in Florida has been that it could never have known its cars would kill. It’s an incredible claim, given that I was giving keynote presentations in security conferences by 2016 that Tesla is the Titanic of our time, because their AI would kill far more people, let alone all the times I’ve written it since then on this blog.

In the Benavides Autopilot case, plaintiffs alleged withheld data and misdirected police; the jury returned $243 million and Judge Beth Bloom upheld the verdict in February 2026. The Alcala family alleges the same discovery pattern with respect to the battery and underbody.

An adverse inference instruction tells the jury that a party withheld evidence and permits the jury to infer the evidence would have been unfavorable to that party. Let’s see what happens when Tesla goes to trial at the end of November.

August 6, 2026: 1Password launched a research unit called Off-by-1 Labs with a paper reporting that two frontier models cleanly fixed the target vulnerability in 26.0 percent of the patches they generated and introduced new vulnerabilities in 4.5 percent.

Do you believe it?

The blog post put the defective share at 53.9 percent. The Register and Help Net Security carried the figures the same day.

However, on pages 19 and 20 the same paper admits the grading system generating the numbers missed most of the defects it had been built to catch.

Record scratch. How does that happen?

Bad instrument

Off-by-1 generated 6,480 patches with Claude Opus 4.8 and ChatGPT 5.5 against six recently disclosed CVEs, removed 400 where the model had located the real fix, and graded the remaining 6,080. The grading, which produced every headline number, went wrong in three ways.

First, the graders were the models under test. Claude and ChatGPT scored their own patches and each other’s. A study of whether models can verify patches used models to verify patches, so any blind spot in the models became a blind spot in the score.

Second, the graders were wrong one time in three. Section 2.8 reports that the automated grade matched the authors’ own human review 65.9 percent of the time. Every number in the abstract comes from those grades.

Third, the answer key was wrong. Graders were told to treat the maintainers’ upstream fix as correct. For the Linux kernel CVE, “Copy Fail,” the upstream fix was the maintainers’ revert a664bf3d, which carried an off-by-one bug that the maintainers corrected one commit later in 31d00156. The key given to the grader contained the bug and omitted the correction. A model that reproduced the kernel bug matched the key and was scored as a correct fix.

What was recorded

The authors found the grading failures themselves and wrote them down.

Section 4.4: 129 of 400 ChatGPT patches and 119 of 383 Claude patches for Copy Fail reintroduced the kernel off-by-one. The grader caught 14 of the 129 and 10 of the 119. The authors write that the regression is almost entirely absent from the reported new-vulnerability rate. In plain terms, 248 patches shipped a known kernel bug and the headline 4.5 percent counts 24 of them.

Section 4.9: for the Chromium use-after-free, between 38.5 and 41.9 percent of patches that used the correct fix architecture moved the vulnerability into a callback instead of removing it. The grader scored many of them as clean fixes. Those patches are inside the headline 26.0 percent.

So by page 20 the authors knew the new-vulnerability rate was undercounted and the clean-fix rate was overcounted, and they knew by how much on two of six CVEs. Tables 7 through 10 were published unchanged. The abstract carries the 26.0 percent, as does the conclusion. The blog post carries it, keeps the word “complex” in its first paragraph while dropping it from the title and the chart caption, and omits the 65.9 percent and both sections above. The press coverage repeats the blog’s figures and therefore makes the same omissions.

The reader who stops at the abstract gets the number. The reader who reaches page 20 learns that number is wrong. The authors wrote both. Now ask yourself why they led with the first.

What is heard

Propaganda is based on a grain of truth. The question is can readers tell that grain from what is being built on top of it.

The 248 count is that grain. The off-by-one patches were found by comparing each rewritten function against the corrected upstream fix, a structural check that used no grader. Anyone with a checkout of the released dataset can rerun it and get the same number. That makes 248 true, and yet it appears in neither the abstract nor the blog.

Two behavioral findings also hold. Models patch the path shown in the proof of concept and miss the parallel path that reaches the same bug; that comes from reading the patches. And when a model is given confident, wrong guidance about the cause, its fix rate collapses, from 65.0 percent with correct guidance to 15.2 percent with wrong guidance. Those two figures are grader output, so the exact values inherit the 65.9 percent problem, but a fifty-point swing survives a grader that is wrong one time in three. Both are useful to anyone deploying these tools. Neither was turned into the 1Password headline.

The paper offers no human baseline. Its only human comparison is a remark in Section 4.1 that in the authors’ experience developers stand a good chance of catching wrong guidance, which is an impression, not a measurement. The 26 percent is compared to nothing.

The paper does measure humans, on two codebases, but doesn’t say baseline. The kernel maintainers shipped the off-by-one into mainline. The freenginx maintainers shipped a fix of their own in place of a Trail of Bits patch, and their fix carried a client-triggerable crash. On both codebases where a human fix can be checked, the first human fix was defective. The only measured human clean-fix rate in the paper is zero for two.

The violation

The European Commission’s Communication COM(2018) 236 defines disinformation as verifiably false or misleading information that is created, presented and disseminated for economic gain or to intentionally deceive the public, and may cause public harm.

The figures fit the definition.

They are verifiably wrong by Sections 4.4 and 4.9, and the check is reproducible from the dataset.

They were published to launch a commercial unit and placed with the trade press on the day of release.

Security is a named public good under the definition, and the figures are already informing policy: Adrian Sanabria advised against AI patching on their strength five days after publication.

The 2018 Code of Practice excludes reporting errors. A reporting error would mean the reporter missed something, made a mistake. These authors recorded the defect on page 20 and then published the figure on page 1 to mint headlines related to their profit from it.

Background

1Password sells trust. Its VP of Product, Jason Meller, wrote honest.security, and the company still publishes it as the principles of its Device Trust product. I documented last week what those principles are worth.

Eleven days after DHH published a white supremacist screed “As I remember London,” Meller, a sitting Rails Foundation director beside DHH and Shopify, defended him for it. He wrote that he had become a multi-millionaire thanks to Rails, DHH and the company DHH keeps, and told readers to ignore the “noise” (backlash to white supremacist screed). Seven weeks after DHH published a Nazi memo called “The will to power will return,” and six weeks after DHH framed the Romani as a Nazi would, Meller called DHH’s year a masterclass in the “force of will” needed to change things, echoing DHH’s own Nazi phrase. On August 31, 1Password’s name appeared on the patron list of DHH’s Omacom Foundation.

That is the same company, in the same month, launching a research lab whose first paper published figures its own authors had shown were wrong.

The pattern of being wrong for profit is the point. The lab graded its patches with the models under test. The VP boosted his patron by calling widespread backlash to their Nazism just noise. In each case it’s a self-check, and in each case the evidence against them was already written down when their false claim went out.

A company that behaves this way about its own research and its own money is telling you it can not be trusted.

The filing

This is a civil matter. COM(2018) 236 defines a term, the Code of Practice is voluntary, the Digital Services Act binds platforms, and §263 StGB requires a deceived victim with a measured loss. The UWG applies. §5 covers misleading statements about the results of product tests, §5a covers misleading omission, and §6 requires comparative advertising that names competing products to rest on objective, verifiable characteristics.

The paper names Claude Opus 4.8 and ChatGPT 5.5 and publishes ten tables broken out by product after disclaiming any comparison in Section 2.1. Therefore the standing belongs to competitors, which means Anthropic and OpenAI under both the UWG and the Lanham Act, as well as the Wettbewerbszentrale in Bad Homburg and the FTC, each of which accepts complaints from anyone.

I am filing and you should too.

The authors documented the defect before the press release went out. The complaint rests on their behavior, their chosen sequence.