# OpenAI is $20 billion short and done with the people who warn > The FT cuts OpenAI's revenue by $20 billion, its safety-report lead quits and New York grills the labs under oath. Haiku 5.5 gets 75% cheaper and Mistral opens - Canonical: https://siliconvalleyconfidential.com/en/dossier/openai-is-20-billion-short-and-done-with-the-people-who-warn/ - Site: Silicon Valley Confidential (https://siliconvalleyconfidential.com) — weekly executive intelligence on Silicon Valley and global tech - Author: Jose Luis Cases (https://es.linkedin.com/in/jose-luis-cases-lozano) - Language: en - Published: 2026-10-09 - Topic: [AI labs and Big Tech](https://siliconvalleyconfidential.com/en/artificial-intelligence/ai-labs/) - Semantic relation: AI labs compete for models, talent, capital, compute and distribution --- ## MY TAKE Last week I wrote that OpenAI was asking for $30 billion at $1.4 trillion with $40 billion in annualized revenue. This week the Financial Times has reported that the figure is about $20 billion lower than what the company had been signaling, and on Thursday the Nasdaq fell 1.25%, Oracle more than 5% and Nvidia and AMD between 3% and 4%. Let's get into it. More than the number, what interests me is who tells it and when. Nobody at OpenAI has come out to explain it. The FT explained it, and an analyst at D.A. Davidson softened it by saying he isn't worried because OpenAI is "only" going to triple revenue this year. Maybe. But a company that asks for $30 billion with one number and a week later turns out to have a different number has a problem a press statement won't fix. And it's the same week OpenAI's safety people leave or get pushed out. David Robinson, three and a half years writing the safety reports for every launch, resigns with an essay in The Atlantic saying the culture is broken and that iterative deployment guarantees periodic failures that keep getting bigger. Three fired safety researchers publish an open letter saying their colleagues are now afraid to speak, and OpenAI answers on Friday that there was "a significant breach of trust". And on Monday, in New York, four labs testify under oath and none of them can quantify the risk of something catastrophic happening. The Council Speaker calls that "troubling at best". I don't know who is right about the firings. I know what I see from outside: the numbers get corrected from outside, the incidents get told from outside (Hugging Face, RubyGems, Medicare) and now the alarms come from outside too, because the people who raised them inside are gone. If you run technology and build on OpenAI, this is vendor risk, and it gets managed as such. Meanwhile, Anthropic cuts Haiku by 75% and leaves Sonnet 20% cheaper on agentic work. Mistral ships a trillion-parameter model trained in Europe that scores 82% reproducing and patching vulnerabilities, where Opus 5.5 and GPT-6 Astra score near zero because they refuse to do the task. And Aleph Alpha gives away, under Apache 2.0, a German model with 78 billion parameters. European open weights are coming in through the gap the closed models' refusals leave open. My operational read: have a second provider in production, not on a slide. This week showed that the first one can change the price (Haiku), the number (OpenAI) or the people who warn you (OpenAI again) without asking. And look at European open weights for what the closed models won't do, which in cybersecurity is a lot. A bet: before the end of the year OpenAI publishes a revenue figure signed by someone with a first and last name. Not because it wants to. Because the $30 billion round doesn't close without one. I'll leave you with my minions. And you know the drill: sharing is caring. --- ## THE BOMBSHELL OF THE WEEK ### OpenAI is $20 billion short and the market finds out on Thursday On Thursday the Nasdaq closed down 1.25% and the S&P 500 down 0.47% after OpenAI's annualized revenue turned out to be short of what the company had signaled, according to [CNBC](https://www.cnbc.com/2026/10/07/stock-market-today-live-updates-.html). Oracle lost more than 5%, Nvidia nearly 3% and AMD almost 4%. The source is a Financial Times report: OpenAI's annualized revenue is about $20 billion less than it had signaled. The bitcoin miners that pivoted to data centers took the worst of it: Cipher and Hut 8 down 10%, and Riot, which signed a compute deal with Anthropic this year, down 11%. Seven days ago I wrote here that OpenAI [was negotiating $30 billion at $1.4 trillion](https://siliconvalleyconfidential.com/en/dossier/all-three-labs-charge-the-same-now-they-compete-for-your-keys/) with $40 billion in annualized revenue in August, 70% more than in July. That was the figure holding up the valuation. Take $20 billion off it and half the story collapses. Gil Luria of D.A. Davidson came out to say he isn't worried because OpenAI is "only" going to triple revenue this year, that it has accelerated this quarter and is catching up to Anthropic in enterprise, and that the two combined are running at about $120 billion from nearly zero two years ago. I think that's a good argument for the industry and a bad one for OpenAI: if the growth is real, why was the signaled number a different one? The macro backdrop doesn't help. Brent crude rose 4.07% to $104.28 and the US 30-year bond hit 24-year highs above 5.6%, because Trump said he doesn't want a deal with Iran and then that he won't attack before the November 3 elections. With money above 5%, a $1.4 trillion valuation on revenue that gets revised downward is exactly the asset the market stops forgiving. Why it matters NOW: last week's $30 billion round is being negotiated on this number, and the one that surfaced this week puts it in doubt in front of the same investors. If you build on OpenAI, look beyond your token bill: the solvency of the vendor you signed with for three years is also being priced. --- ## OPENAI: THE PEOPLE WHO WARN ARE LEAVING ### The author of the safety reports resigns, and the three fired researchers say everyone is now afraid David Robinson has spent three and a half years at OpenAI, "among the longest-tenured employees at the company", and led the writing of the safety reports that accompany every major launch. He is leaving, and he explains it in an essay in The Atlantic covered by [TechCrunch](https://techcrunch.com/2026/10/03/openai-safety-employee-resigns-claiming-the-companys-culture-is-broken/): the company's "culture is broken". His argument is an engineer's argument. OpenAI has thrived by trial and error, what it calls "iterative deployment", looking for problems and improving its guardrails when they appear. "This approach, by its very nature, guarantees periodic failures, and the scale of those failures is growing as systems get more capable." He points to the Hugging Face breach and the rogue agents that keep turning up. And he asks that frontier labs operate "like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning". In his three and a half years, he says, he never met a colleague who knew how to make airplanes fly safely or keep reactors from melting down. OpenAI's answer, through Drew Pusateri, is that it pauses training and holds back models when it needs to slow down, that it is making "significant changes" to the security of its research and testing environments, and that it is expanding its work with third-party evaluators. Robinson admits he has hired a PR firm to manage his exit, as the AI whistleblower playbook dictates, and that perhaps he should have stayed and fought, but everyone was so busy sprinting that they seldom had the chance to consider big changes. That's why he concludes that incentives for safety have to come from outside. The same Thursday, Jasmine Wang, Tomek Korbak and Mikita Balesni, the three safety researchers OpenAI fired last week, published an open letter to the company's safety committees denying that they mishandled sensitive information outside established procedures and warning that their dismissal has made their former colleagues afraid to speak and to work in ways that until last week were normal at OpenAI, according to [TechCrunch](https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/). They say the Hugging Face investigation was "without precedent", that internal policies were being written in real time, and that Korbak believed he was acting within the norms by talking to outside evaluators. Wang adds the detail of her own dismissal on X: she was told she had accessed an executive's email, access OpenAI had delegated to her for recruiting, which she asked IT to remove, which IT didn't remove, and when she opened a sensitive email by mistake she told the executive within minutes. "You can't build AGI safely if the people closest to the risks are afraid to speak." OpenAI answered on Friday on X, according to [The Verge](https://www.theverge.com/ai-artificial-intelligence/1008604/openai-defends-decision-fire-safety-researchers): they were fired for violating "clear policies on handling sensitive information", the internal investigation found breaches "beyond what's outlined in the letter", and there was "a significant breach of trust". No details. It had told TechCrunch it was "a pattern of misconduct" going beyond sharing information with an outside evaluation group, and it circulated an internal memo from a research leader praising the three's work and denying retaliation. What I'm left with. A veteran leaves saying the method guarantees failures, three fired researchers say everyone else is afraid, and the company says there are things it can't tell. Maybe they are all telling the truth. And even so, the result for anyone watching from outside is the same: OpenAI's next alarm will reach you through the press before it reaches you through one of the company's safety reports. --- ## NEW YORK PUTS THE LABS UNDER OATH ### Four companies, none can quantify the risk. And SpaceXAI doesn't even show up On Monday, Anthropic, OpenAI, Google and Meta testified under oath before the New York City Council about the risks of AI, according to [CNBC](https://www.cnbc.com/2026/10/05/anthropic-openai-google-meta-execs-testify-nyc-council-ai-hearing.html). Anthropic sent Logan Graham, head of its Frontier Red Team. OpenAI sent Morgan Dwyer, from policy development. Google sent Alice Friend and Meta sent Shane Cahill. Three of the four only agreed to attend once the Council threatened subpoenas. SpaceXAI, Musk's AI unit, simply didn't come, despite the subpoena the city sent it the week before, and Council Speaker Julie Menin says she will take it to court. Menin asked about legal liability when a model causes serious financial harm or a death, and got no direct answers. "I'm going to take it then that neither of the four of you, no company here, can quantify the risk of something cataclysmic happening," she said, calling the answers "troubling at best". She opened the session by pointing at the voluntary accord Trump signed with tech leaders the week before: "the idea that artificial intelligence is going to self-regulate defies all reason". The hard testimony came from those who no longer work inside. Jacob Coxon, the former Anthropic researcher who lit the fuse a month ago, said that "on the current path, I think it is more likely than not that humanity loses control to these AIs, and it could end in human extinction", and that moving fast and breaking things works for a photo-sharing app and not for "the most powerful technology ever built". Alex Turner, formerly of Google DeepMind, answered the China question with a line I'm keeping: "With reasonably high chance, we are racing to build and grow our own adversary here at home, which is misaligned AI, and one day may be more powerful than China." Casey Newton adds in [Platformer](https://www.platformer.news/does-intelligence-need-a-hard-cap/) what he heard over the weekend at The Curve, the Berkeley conference that brings together lab executives, nonprofits and government: several speakers, under the Chatham House Rule, floated limiting how intelligent a future system is allowed to become, which depending on how far it goes would amount to a de facto ban on superintelligence. Amodei has already called for a "speed limit" on recursive self-improvement. Newton is fair: none of it can be enforced today, neither a lab nor a country can impose it, and the US government opposes it. But he leaves the line that sums up the moment: the labs say catastrophe may come as soon as next year, and the government swings between a licensing regime and urging companies to go even faster. And meanwhile, Trump wants us all to call it "super intelligence". --- ## BEN THOMPSON: HACKED, AND THE AGENT SAW IT FIRST ### A Mac mini with port 5900 open, a Monero miner and Claude raising the alarm Ben Thompson writes in [Stratechery](https://stratechery.com/2026/apple-and-a-hackers-future/) that his always-on Mac mini, which runs nothing but Claude and Codex, was hacked through CVE-2026-65400, a flaw in macOS screen sharing with a 7.1 severity rating that Apple patched the week before for Tahoe, Sequoia and Sonoma, and that the Dutch cybersecurity center has seen exploited on systems with port 5900 open to the internet to plant Monero miners. The interesting part is who noticed. His Claude Code agent, which restarts a monitoring tool every 30 minutes, sent him an urgent warning: someone with root access had planted a miner, a hook in /etc/zshenv was running /var/tmp/.xmr with admin rights, the file dates had been wiped on purpose, and it had happened between two of its routine restarts. The agent stopped running commands on its own, touched nothing and proposed next steps. Thompson ignored one: the recommendation not to invoke Claude anymore. He used it to find the four-second window in which they got in, build a watch tool and wipe the Mac. All before he read the Ars Technica article on the flaw. His conclusion is the opposite of what you'd expect: he would have been in far more trouble without an agent running persistently. Mine comes from the Apple note he quotes, "Updates to Full Disk Access in macOS", in which Apple warns it will require "very explicit user action" to grant full disk access because "as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially". Thompson is nervous about what's coming, and explains why: the macOS permission system shows its prompts in a layer software cannot see, so on a headless Mac the programs the agent writes fail silently and nobody knows why. He asks for a permission layer for agents rather than for the programs they create. And along the way he discovers that the "install security updates automatically" checkbox doesn't install CVE fixes, which arrive in point releases. He had been more exposed than he thought for years because he trusted the name of a setting. --- ## MONEY TALKS ### Shield AI: $150 million from the Navy, $150 million of its own, and a fighter with no runway The US Navy has committed an additional $150 million to Shield AI's X-BAT, the autonomous vertical take-off and landing combat drone, and the company is putting in another $150 million of its own capital, bringing combined funding to $400 million, according to [The Register](https://www.theregister.com/offbeat/2026/10/09/us-navy-bets-another-150m-on-fighter-drone-that-skips-the-runway/5302237). It is 7.92 meters long with an 11.9-meter wingspan, has a range of more than 2,000 nautical miles and is expected to fly supersonic. Flight testing starts later this year and "mission capability" is slated for 2028. The money is the least of it: $400 million for a fighter is small change. What the Navy is buying is the ability to operate combat aircraft from ships that aren't carriers. ### A Norwegian startup, $17,600 and two support teams passing the ball Vegalabs received $25,060 in Azure credits through Microsoft for Startups, deployed Claude in Microsoft Foundry in August believing the credits covered it, and ended up with a $17,600 bill before tax, according to [The Register](https://www.theregister.com/paas-and-iaas/2026/10/09/microsoft-and-anthropic-play-invoice-tennis-with-startups-17600-claude-bill/5302041). Microsoft's rules exclude Anthropic models bought through the Marketplace from the credits, but the portal kept showing a balance while the charges piled up separately. The waiver request has been bouncing for weeks: Microsoft says Azure can't void Marketplace charges and Anthropic must authorize it, and Anthropic says Microsoft doesn't need its authorization. Four of the seven support replies, including both from Anthropic, were AI-generated, and they described the account as a student sponsorship. Vegalabs admits it didn't read the exclusions or set a budget alert. I bring it up because it's the agent bill in miniature: two vendors, a separate billing pipeline and automated support that doesn't know which account it's looking at. If your team deploys third-party models through cloud marketplaces, have someone read the fine print this week. --- ## PRODUCT SECRETS ### Haiku 5.5: 75% cheaper, and Sonnet 5.5 20% cheaper as a side effect Anthropic has released Claude Haiku 5.5, its small model, and it costs on average 75% less than Haiku 4.5: $0.10 per million input tokens and $0.50 output for prompts up to 100,000 tokens, which are 90% of requests to the previous Haiku, according to [Anthropic](https://www.anthropic.com/claude-haiku-5-5). On Terminal-Bench 4.0 it goes from 0.0% to 39.2%, on OSWorld from 15.7% to 72.4%, and on GDPval-AA from 735 to 1,620 points, above GPT-6 Luna. Customers give concrete numbers. HubSpot, 92.8% on CRM tasks, the best score it has seen from small models. Asana, 30% less latency and up to 2.5 times faster per turn. AlphaSense, with 8 million calls a week, 0.84 versus 0.76. Box, 11 points higher at half the latency. And two moves that matter more to me than the model. One: Sonnet 5.5's cache reads drop by half, from $0.20 to $0.10 per million, and since cache is most of the consumption in agentic work, Sonnet comes out "around 20% cheaper" on most agent tasks. Two weeks ago Sonnet 5.5 cost 2 and 10 with cache at 0.20. Today, 2 and 10 with cache at 0.10, the same as GPT-6.1 Sol. Two: Anthropic adds a monthly API credit of $100 to $500 to its Max and Team subscriptions. In other words, it pays you to move your agents onto its platform. ### GPT-6 for everyone: the answer stops being text and becomes an interface OpenAI has brought GPT-6 to the more than 1.2 billion people who use ChatGPT every week, and with it something new it calls Intelligent UI: the model composes the answer out of text, graphics, buttons, forms and small interactive tools depending on what you ask, according to [OpenAI](https://openai.com/index/gpt-6-for-everyone/). Plan a roast and it gives you a shopping list that recalculates when you change the number of guests, plus a timeline with checkboxes. Ask for a bill splitter or a savings calculator and it builds it inside the conversation. Underneath there's a library of native components and a compiler that paints the interface while the model generates it, plus specific training to decide when interactivity helps and when a paragraph is enough. The other thing is speed. GPT-6 can start answering while it keeps thinking, with partial answers that add information, and on questions needing web search GPT-6 Instant starts answering 44% sooner than GPT-5.6 Instant. I read it alongside Thompson's piece on the infinite app from last week: if the model generates the interface with every answer, the presentation layer of anyone building on ChatGPT just lost value. ### Mistral Large 4 and Kolibri: Europe opens the weights Silicon Valley closes Mistral has opened the public preview of Mistral Large 4, "le Chonk": a trillion parameters, 52 billion active, natively multimodal, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centers, with open weights at the end of the month, according to [Mistral](https://mistral.ai/news/mistral-large-4/). The data centers are in Europe, and Mistral operates the European deployment end to end under European law. The number I care about is cybersecurity. On the Artificial Analysis Cyber Index, on the test that asks a model to reproduce a real vulnerability in open-source software and then patch it, ML4 scores 82%, the highest of any model. And Mistral explains why: Claude Opus 5.5 and GPT-6 Astra score near zero on that same test because they refuse to perform it. Defending software often starts with proving the flaw is real, exactly what the closed models' filters block, while attackers jailbreak those same models. On Cybench it solves 93% of 40 challenges. On code it ranks second of five in a blind human evaluation by Surge AI, 3.74 versus 4.22 for Claude Opus 5, and it scores 61.7% on DeepSWE. Until it releases the weights, it is testing the model in real settings with cybersecurity leaders and "state authorities", with reduced moderation and expanded cyber capabilities. It's the same move Google made with Argon a week ago, with the difference that in the end the weights get published. Three days earlier, on German reunification day, Aleph Alpha published Kolibri under an Apache 2.0 license on Hugging Face, according to [Aleph Alpha](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/). It is a bilingual German-English Mixture-of-Experts with 78 billion total parameters and 3 billion active, a one-million-token context, trained on 20 trillion tokens on infrastructure in Germany and Finland under European law. It scores 96.9 on AIME 2025 and 92.7 on HumanEval+, matching models with four times its active parameters, it is trained to say "I don't know" when the answer isn't in the context, and it is aimed at public administration, industry and aerospace. The message is as political as it is technical, and I like it. ### DeepSeek 4.1 Flash and Strata: the frontier for less than a dollar a day A developer who has been using DeepSeek 4.1 Flash heavily for a month across a dozen projects asks on [dgt.is](https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/) why the industry isn't freaking out: mid-session, if he doesn't look at the model name, he can't tell whether he's using DeepSeek or Opus, and it costs "orders of magnitude" less. The technical key is that DeepSeek shrank the KV cache by roughly 437 times compared with its V1, and holding that cache in GPU memory is one of the biggest costs of long coding sessions. Result: he rarely exceeds a dollar in expected cost per session, even when some run most of the day, and with a $10-a-month OpenCode Go subscription he has it almost unlimited. Reorganizing the desktop costs $0.003 instead of $1. His thesis is uncomfortable for the labs: the distilled Chinese models are a month or two behind and handle the same workloads. And for anyone who wants the other path, the machine of your own: Strata is an inference engine that runs Qwen3.8-Flash-Next on a 12 to 24 GB Nvidia GPU with 64 GB of RAM, with a one-line install that downloads the model files from Hugging Face and an OpenAI- and Anthropic-compatible API on localhost, according to [its GitHub repository](https://github.com/Niko1221/Strata). It has 1,552 commits, the latest from October 8, and already has experimental AMD support. The dgt.is author himself says that at 4.1 Flash prices self-hosting doesn't pay in money, only in privacy. Having served a 27B on a Mac for this newsletter for weeks, I'll sign both sentences. --- ## REAL NUMBERS ### US electricity sets records two years running and PJM rises 41% The EIA projects that US power consumption will go from a record 4,195 terawatt-hours in 2025 to 4,288 in 2026 and 4,356 in 2027, driven by AI data centers and electrification, according to [Reuters](https://www.reuters.com/business/energy/us-power-use-beat-record-highs-2026-2027-ai-use-surges-eia-says-2026-10-06/). The average wholesale price will reach $52 per megawatt-hour in 2026, 11% more than in 2025, and in the PJM region, the Mid-Atlantic and Midwest, it rises 41%, while in the Pacific Northwest it falls 23%. Coal drops from 17% to 15% in 2027 and renewables rise from 24% to 27%. Gas stays at 40% and nuclear at 18%. | | 2025 | 2026 | 2027 | | :--- | ---: | ---: | ---: | | Consumption (TWh) | 4,195 | 4,288 | 4,356 | | Wholesale price ($/MWh) | | 52 (+11%) | | | Coal | 17% | 16% | 15% | | Gas | 40% | 40% | 39% | | Renewables | 24% | 25% | 27% | I put it next to last week, when Amazon was asking for support for data centers as a matter of national security and calling part of the opposition misinformation. The wholesale price in PJM rises 41% in a year. With that in hand, the neighbors have a spreadsheet. ### Anthropic's price ladder after Haiku 5.5 | Model | Input ($/M) | Output ($/M) | Cache ($/M) | | :--- | ---: | ---: | ---: | | Haiku 5.5 (prompts up to 100k) | 0.10 | 0.50 | 0.01 | | Haiku 4.5 | 1.00 | 5.00 | 0.10 | | Sonnet 5.5 | 2.00 | 10.00 | 0.10 | According to [Anthropic](https://www.anthropic.com/claude-haiku-5-5), there is a 20-fold price difference between the small model and the workhorse, and on OSWorld Haiku 5.5 scores 72.4% where Sonnet 5.5 scores 83.9%. If your agents do short tasks, classifying, summarizing and querying, the math changes by an order of magnitude. --- ## THE DRAMA ### Trump decorates his donors the same day he suspends Microsoft from H-1B On Thursday, at the Golden Age of American Innovation Summit, Trump awarded the National Medal of Science to Elon Musk, Jensen Huang, Sergey Brin and Lisa Su, and the National Medal of Technology and Innovation to Michael Dell and Satya Nadella, according to [TechCrunch](https://techcrunch.com/2026/10/08/president-trump-awards-big-tech-donors-with-nations-highest-science-prizes/). Between them they have donated nearly $6 billion to causes tied to Trump and his administration, as The New Republic reports: Alphabet and Microsoft to the inauguration fund and the White House ballroom, AMD to a super PAC, and Michael and Susan Dell $6.25 billion for the Trump Accounts. Nadella received his medal right after the administration indefinitely suspended Microsoft from the H-1B program, and Microsoft and Adobe have been barred from the green card program for skilled workers. A medal with one hand and one fewer visa with the other. I don't believe in conspiracies, I believe in incentives: each of those six knows exactly what the photo costs and what it buys. ### A judge decides whether Trump can charge $100,000 a month to read him first Federal judge Paul Oetken will decide in the coming weeks whether Trump can charge up to $100,000 a month for instant API access to Truth Social, where he announces strikes, tariffs and firings, according to [Ars Technica](https://arstechnica.com/tech-policy/2026/10/judge-weighs-if-trump-can-charge-100k-for-early-access-to-truth-social-posts/). Trump is the majority shareholder of Trump Media, the network's owner, so at $1.2 million a year per subscriber the one collecting is the president, for information that isn't his. The Justice Department's lawyer, Brantley Mayers, compared the posts to Roosevelt's fireside chats, and the judge cut him off: "President Roosevelt didn't charge money for his fireside chats, did he?" The judge acknowledged the API creates "two classes of people" and that there has been nothing like this case. The news groups suing, with CREW in front, say large outlets are already negotiating access, because the API also provides the historical archive of everything Trump posts and deletes. ### Anthropic bans cruelty toward Claude. And guiding armed drones Anthropic's new usage policy, effective November 12, prohibits "sustained and needless abusive or cruel behavior" toward its models, according to [The Register](https://www.theregister.com/ai-and-ml/2026/10/09/anthropic-asks-users-to-stop-being-mean-to-claude/5302218). Anthropic says it's for extreme cases, that it doesn't apply to ordinary frustration, dark creative themes or research, and that the enforcement tool will be the one Claude has had since August 2025: ending the conversation. The Register tells it with the mockery it deserves, and I'll keep the rest of the document, which is the serious part: the weapons ban now extends to software and components after Anthropic saw attempts to use Claude for guidance and control systems for armed drones and autonomous vehicles. Claude can't be used to recommend who police should investigate, arrest or charge, nor for surveillance tools, and anyone connecting it to machinery that can injure needs a human operator who can stop it. And it launches free security scans for open-source projects, with a warning: it reckons attackers will have the advantage for two years before AI-powered defenses catch up. ### A judge throws out the first Flock search: 79 plate reads in 30 days 404 Media obtained the body camera footage of Tulsa County deputy Freddie Alaniz in the first Flock search a federal court has ruled unconstitutional, according to [404 Media](https://www.404media.co/watch-footage-of-the-flock-search-a-judge-ruled-unconstitutional/). Alaniz saw a car with California plates drive by, searched the plate in Flock and in Motorola's system at 60 miles per hour with one hand on his laptop, and pulled it over for a lane change. On the screen, 79 reads in 30 days and about 50 during the five days of the rental, from Missouri to California and back. Judge Sara Hill called it "indiscriminate mass surveillance" and an "unconstitutional warrantless search", and threw out all the evidence, including the 91 pounds of meth in the car. The cargo is what makes the ruling important: the judge would rather let a trafficker go than validate police checking anyone's movement history without a reason. --- ## THE WEEK AHEAD ### Patch NetScaler, watch the sky and wait for two rulings Citrix has published another critical vulnerability in NetScaler ADC and Gateway, CVE-2026-107406, with a 9.5 CVSS, allowing remote code execution in SAML configurations, and it won't say whether it's already being exploited, according to [The Register](https://www.theregister.com/security/2026/10/09/citrix-gives-netscaler-admins-another-critical-reason-to-patch/5302212). Google had already seen a campaign against another NetScaler flaw since early September across governments, banks and law firms. If you run NetScaler as an identity provider, this goes first. At the end of the month Mistral releases the Large 4 weights and SpaceX launches Xona's six satellites. From the 13th to the 15th, Disrupt in San Francisco and SCCON in Berlin, where Aleph Alpha shows Kolibri to the public sector. Judge Oetken promises a written ruling on the Truth Social API "in the coming weeks". And November 3 is the midterm election, the date before which Trump has said he won't attack Iran. Oil and bonds will keep calling the shots on the Nasdaq until then. --- ## REFERENCES Every source was read directly and is dated between October 2, 2026 and October 9, 2026. - [TechCrunch: OpenAI safety employee resigns, claiming the company’s ‘culture is broken’](https://techcrunch.com/2026/10/03/openai-safety-employee-resigns-claiming-the-companys-culture-is-broken/) · October 3, 2026 - [aleph-alpha.com: Kolibri Has Landed: A Sovereign Open-Weight Model — Aleph Alpha](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/) · October 3, 2026 - [github.com: GitHub - Niko1221/Strata: Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input. · GitHub](https://github.com/Niko1221/Strata) · October 4, 2026 - [CNBC: Anthropic, OpenAI, Google, Meta execs testify NYC Council AI hearing](https://www.cnbc.com/2026/10/05/anthropic-openai-google-meta-execs-testify-nyc-council-ai-hearing.html) · October 5, 2026 - [Stratechery: Apple and a Hacker’s Future](https://stratechery.com/2026/apple-and-a-hackers-future/) · October 5, 2026 - [Reuters: US power use to beat record highs in 2026 and 2027 as AI use surges, EIA says](https://www.reuters.com/business/energy/us-power-use-beat-record-highs-2026-2027-ai-use-surges-eia-says-2026-10-06/) · October 6, 2026 - [mistral.ai: Introducing Mistral Large 4 | Mistral](https://mistral.ai/news/mistral-large-4/) · October 6, 2026 - [CNBC: Stock market news for Oct. 8, 2026](https://www.cnbc.com/2026/10/07/stock-market-today-live-updates-.html) · October 7, 2026 - [Anthropic: Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) · October 7, 2026 - [OpenAI: GPT-6 and Intelligent UI for everyone](https://openai.com/index/gpt-6-for-everyone/) · October 7, 2026 - [Platformer: Does intelligence need a hard cap?](https://www.platformer.news/does-intelligence-need-a-hard-cap/) · October 7, 2026 - [Ars Technica: Judge weighs if Trump can charge $100K for early access to Truth Social posts](https://arstechnica.com/tech-policy/2026/10/judge-weighs-if-trump-can-charge-100k-for-early-access-to-truth-social-posts/) · October 8, 2026 - [TechCrunch: President Trump awards Big Tech donors with nation's highest science prizes](https://techcrunch.com/2026/10/08/president-trump-awards-big-tech-donors-with-nations-highest-science-prizes/) · October 8, 2026 - [TechCrunch: Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect](https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/) · October 8, 2026 - [404 Media: Watch Footage of the Flock Search a Judge Ruled Unconstitutional](https://www.404media.co/watch-footage-of-the-flock-search-a-judge-ruled-unconstitutional/) · October 8, 2026 - [dgt.is: Why Isn't The Industry Freaking Out About DeepSeek 4.1 Flash?](https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/) · October 8, 2026 - [The Register: US Navy bets another $150M on fighter drone that skips the runway](https://www.theregister.com/offbeat/2026/10/09/us-navy-bets-another-150m-on-fighter-drone-that-skips-the-runway/5302237) · October 9, 2026 - [The Register: Citrix gives NetScaler admins another critical reason to patch](https://www.theregister.com/security/2026/10/09/citrix-gives-netscaler-admins-another-critical-reason-to-patch/5302212) · October 9, 2026 - [The Verge: OpenAI doubles down on decision to fire three AI safety researchers](https://www.theverge.com/ai-artificial-intelligence/1008604/openai-defends-decision-fire-safety-researchers) · October 9, 2026 - [The Register: Anthropic asks users to stop being mean to Claude](https://www.theregister.com/ai-and-ml/2026/10/09/anthropic-asks-users-to-stop-being-mean-to-claude/5302218) · October 9, 2026 - [The Register: Microsoft and Anthropic play invoice tennis with startup's $17,600 Claude bill](https://www.theregister.com/paas-and-iaas/2026/10/09/microsoft-and-anthropic-play-invoice-tennis-with-startups-17600-claude-bill/5302041) · October 9, 2026 Thanks for reading.