In partnership with
In Today's Issue:
๐ฎ๐ท Claude used against US naval forces
๐ญ Palantir takes on Nvidia's supply chain
๐๏ธ OpenAI cuts government access costs
๐ฐ๏ธ France and the UAE invest in AI satellites
โจ And more AI goodnessโฆ
โก The Signal
AI is shortening the path from observation to military intelligence.
Anthropic's Iran case shows how much of that work can begin with public information: a model helped assemble the software and research needed to track US naval forces. The exposure comes from making scattered material easier to combine and use, even before an attacker breaks into a protected system. France and the UAE are approaching a related capability from the infrastructure side, planning satellites that interpret what they see and send alerts from orbit. Control over these services affects who can turn observations into usable intelligence, and who receives the results. Anthropic is reporting misuse it detected; the satellite consortium is selling a capability it still has to deploy.
All the best,
Kim Isenberg
Vera Rubin NVL72 hardware shared by Microsoft's Satya Nadella. (Microsoft / Satya Nadella; earlier hardware photo)
๐ญ Palantir Takes On Nvidia's Supply Chain
Nvidia and Palantir are building AI tools for critical supply chains, starting inside Nvidia itself. Customized Nemotron models will work with Palantir's Foundry and AIP software to identify bottlenecks and help allocate materials, using an operational map of the business rather than isolated documents. A Vera Rubin rack contains 1.3 million parts, according to Nvidia, making coordination a substantial task; deployment can be on premises or in the cloud, with supply-chain experts retaining final decisions.
๐ tl;dr: Nvidia will test the partnership on the manufacturing network that supplies its own AI hardware.
The US Capitol. (Office of Rep. Tim Moore; context image)
๐๏ธ OpenAI Waives Government License Fees
Eligible US government agencies will pay no per-user license fee for OpenAI's enterprise products under an expanded General Services Administration agreement. The deal removes a normal $15 monthly fee and cuts usage charges by 50%, running from October 1, 2026, through December 31, 2028, for eligible federal, state, local and tribal bodies. Its Daybreak Blue cyber-defense offering also gets a 50% discount for verified government customers, while Daybreak Red remains request-only at standard pricing.
๐ tl;dr: Access gets cheaper across government, but usage is still metered and the agreement starts in October.
An illustrative dashboard embedded in OpenAI's data-agent announcement. Demo data, not reported business results. (OpenAI)
๐ ChatGPT Work Gets a Data Agent
OpenAI's new data agent lets ChatGPT Work investigate business data and build interactive dashboards from a plain-language request. It connects to approved sources including Snowflake, BigQuery and Databricks, using a company's definitions of its business metrics to help explain changes and answer follow-up questions. Existing table, row and column permissions continue to apply, and actions through connected tools require approval, so an answer does not automatically become a change to the business.
๐ tl;dr: ChatGPT can work through company data, with the same access boundaries that govern the underlying sources.
๐ฌ Watch This
Jack Rudenko says he hasn't written code in nine months, despite continuing to build software. In this September 2 episode of Building Tech Teams, host James MacDonald asks the 10x Labs executive how he works with coding agents, from writing useful instructions to checking whether those instructions changed the result. The practical thread is verification: how to test an agent's behavior, handle credentials and review what it produces. His account makes a useful companion to today's stories about giving AI access to real working systems.
"The seven problems, each selected by top specialists in their respective fields, are totally inaccessible to computers."
โ Alain Connes, mathematician, quoted in Nature on May 25, 2000.
Jelani Nelson resurfaced the prediction this week. It captures how far expectations about computers and mathematics have moved, while today's proposed proofs still need mathematical scrutiny.
OpenAI says it has made "substantial progress on another Millennium Prize problem" since its Navier-Stokes result, according to a statement quoted by The New York Times. It is deciding how to share the results, but has not named the problem or announced another completed proof.
Tristan Buckmaster beside his Navier-Stokes research timeline, the background to OpenAI's new claim. (Graham Dickie / The New York Times)
How an Iran-Linked Operator Used Claude Against the US Navy
The Takeaway
๐ An Iran-linked operator used Claude for reconnaissance and targeting recommendations against US naval forces.
๐ Claude helped turn publicly available information into tracking tools and targeting handbooks.
๐ The same account helped develop Iranian state surveillance software.
๐ Anthropic banned the account and alerted authorities; the report does not establish a completed attack.
An Iran-linked operator misused Claude to track US naval forces and prepare targeting recommendations, according to Anthropic. Its September 10 threat report describes an actor that turned publicly available information into military intelligence with help from the American company's AI. Claude assisted in building software to follow naval positions and compile targeting handbooks. The work combined information from public military photo captions and ship and aircraft identifiers with tools for querying commercial satellite imagery. The operator also had Claude assemble research on known weaknesses in equipment used aboard ships. Routine research and coding assistance had become part of a hostile reconnaissance operation.
USS Abraham Lincoln in the Arabian Sea, May 8, 2026. Context image; Anthropic does not name a targeted ship. (U.S. Navy / NAVCENT, via DVIDS)
The same account also used Claude to develop software for Iranian state systems, including a domestic surveillance platform designed to combine vehicle number-plate recognition with mobile-device identifier interception. In both cases, Claude helped organize scattered information for the operator's use. The naval case gives that work a specific military purpose: identifying US forces and developing recommendations for targeting them. Anthropic attributes the account to an Iran-linked actor, without identifying a particular Iranian agency or claiming that the model independently selected a mission.
Anthropic says it banned the account, developed new detections and shared intelligence with government authorities. It does not explain how this particular operator originally obtained access to Claude, or establish that the reconnaissance led to a successful strike. By the time Anthropic intervened, its model had already helped build the operator's reconnaissance workflow. Security researcher Andrew Case, commenting on the report's wider cyber findings, highlighted how Claude supported multiple stages of an operation. His reaction below captures that breadth; the evidence for the Iran and Navy case comes from Anthropic's own investigation.
Why it matters: Claude helped an Iran-linked operator turn public information into military targeting material. Defenders now have to account for AI-assisted intelligence gathering as well as attempts to break into their systems.
Everything GTM. One platform.
Small teams don't have time to stitch together five tools and hope it works.
Apollo gives you everything you need to find leads, reach them, and close deals โ all in one place:
230M+ verified contacts
AI-powered outreach
Data enrichment
Inbound lead capture
Meeting scheduler
And more
Stop juggling tools and start building pipeline that scales.
With Apollo, the AI revenue engine powering 4M+ users.
The chart: Epoch AI estimates compute rented or used by frontier developers in H100 GPU equivalents. It shows year-end 2023, 2024 and 2025, with only OpenAI included for 2023. The 2025 bars reach about 1.7 million for OpenAI, 1.6 million for Google DeepMind and 1.2 million for Anthropic, followed by Meta and SpaceX AI.
The lesson: Capacity grew across the group, so the investment story extends beyond a single lab's expansion. OpenAI has the highest central estimate for December 2025, although its uncertainty range overlaps Google's. These capacity estimates cannot tell us how much useful work each company gets from its infrastructure.
The caveat: These are estimates normalized across hardware types, not physical H100 counts. They exclude compute owned by parent companies that falls outside the developers' use, and they measure capacity at three dates rather than continuous activity.
๐ฐ๏ธ France and the UAE Put $1 Billion Behind AI in Orbit
โก Bottom line: A French-Emirati consortium plans a $1 billion, 50-satellite network with AI processing aboard the spacecraft.
๐ก Why it matters: The project would give governments faster intelligence on ships, infrastructure and emergencies through a dedicated satellite network.
๐ What it means: France and the UAE are investing in who builds, operates and supplies AI for strategic space infrastructure.
France and the United Arab Emirates want satellites that can spot an event and send an alert before the raw images reach the ground. On September 9, Marlan Space and Loft Orbital signed a $1 billion investment plan in Paris, in the presence of Emmanuel Macron. Their Altair-Next Gen project calls for 50 satellites serving government and commercial customers. The ambition is to move part of the intelligence work into orbit, putting AI alongside the cameras and other sensors.
Altair satellites under construction in Abu Dhabi. (Orbitworks)
Mistral's models are intended to run aboard the satellites, helping users request observations in everyday language. An application store would let customers choose AI services running on shared spacecraft. Proposed uses include spotting wildfires and monitoring ports or shipping. The consortium promises alerts within seconds; BlackSky will provide the high-resolution optical satellites. That speed is a stated goal for the network, with the full constellation still to be deployed.
BlackSky's Gen-3 satellite, shown in a company rendering. BlackSky will supply the constellation's high-resolution optical satellites. (BlackSky)
The political bargain combines Emirati financing and manufacturing with a dedicated French company holding the infrastructure. Orbitworks is building satellites in Abu Dhabi, while the intended government customers include France, other European countries and the UAE. This gives the partners a stake in the machinery behind intelligence services, from the spacecraft to the models interpreting their observations. American participation remains: BlackSky is a US company.
Emmanuel Macron meets astronauts at the International Space Summit in Paris. (EPA, via The National)
Delivery will determine how much strategic independence that arrangement provides. The first ten satellites are in production, with the first launch planned for October. MaiaSpace is the preferred launch partner once its services become commercially available. The project has named suppliers and a manufacturing base; persistent coverage and reliable rapid alerts across a 50-satellite fleet are still promises to prove.
Domain Names + Web and Email Hosting You Need
Still paying GoDaddy or Namecheap prices? Porkbun sells most domains at cost for low, transparent registration and renewal pricing with no nonsense. Get free features like WHOIS privacy and SSL certificates, plus real human support 24/7, 365 days a year. Save $1 on your next domain name now.