Ohio towns vote for data center limits, and Brown wins Senate seat
Ohio held 21 data center ballot measures across 18 municipalities in the 3 November US midterms. That's more than half of the 41 such measures on ballots nationwide. Fortune counted at least 14 Ohio places that passed bans or other restrictions. In Sunbury, unofficial results show a charter amendment passing that bans data centers using 25 MW or more, aimed at a possible Amazon hyperscale project. Trenton approved a ban by about 80% to 20%. In Wilmington, about 80% of voters rejected rezoning in four referendums.
Democrat Sherrod Brown beat Republican Senator Jon Husted for an Ohio Senate seat. Brown wants to end data center tax breaks and make operators pay for the utilities their campuses need. Husted supported the tax breaks.
For builders: Expect new US capacity to take longer to site and carry more local conditions, and expect energy costs to show up in compute pricing. Running workloads across several regions or providers reduces that risk.
OpenAI staff raised security concerns for years, Reuters says
On 3 November Reuters published an account from five unnamed sources. They say OpenAI employees spent several years telling leadership and the board that internal security was weak, and some of those complaints predate ChatGPT's late 2022 launch. A central worry was the company's single shared code repository, which gave anyone with access a view of unreleased projects. Three sources said internal tools, including the sandboxes researchers use to test models, didn't get routine security reviews, and changes to them weren't consistently checked. Some staff felt executives cared more about keeping up with Google and Anthropic.
OpenAI rejects parts of this. It says its most sensitive projects live in separate, restricted repositories, model weights aren't stored in the monorepo, and internal tools now get regular checks under security chief Dane Stuckey. Two sources said things have started to change: in August, a quarter of production engineers were moved onto security. Separately, a covert recording device turned up in a room senior executives used, and prosecutors have charged a former engineer.
For builders: When you review a model vendor, look past API security and ask how they handle insider risk and internal access.
OpenAI partners with Cathedral and raises bug bounty payouts
On Monday 2 November OpenAI announced a new security sprint. It's working with Cathedral, a startup founded by former DOGE staff, whose team "will use OpenAI's models alongside our teams to challenge our defenses." OpenAI also said it's "more than quadrupling our rewards for critical findings" in its bug bounty program.
Cathedral broke into parts of Anthropic's private systems during Anthropic's bug bounty program. Logan Graham, Anthropic's cybersecurity lead, called the method "unusual and out of scope" because it chained several third-party vulnerabilities. He said model weights and Claude conversations weren't affected.
For builders: The big labs now use AI models to attack their own systems. If you run a bounty program, write tight scope rules, including how you treat third-party flaws.
Apple opens Siri AI to developers and flags agent risk to investors
Apple released a Siri AI developer preview on 3 November in the iOS, iPadOS, macOS and visionOS 27.3 betas. Apps plug in their content and actions through App Intents, and a new feature called Skills lets Siri AI combine several actions to handle a complex request. Customers get it in a few months, and App Store submissions open in spring. It isn't available in the EU on iOS, iPadOS or watchOS.
Apple's annual filing, reported by CNBC the same day, warns that agents "may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized." It adds that Apple's "ability to oversee and control third-party artificial intelligence agents" may be limited.
For builders: iOS teams have a few months to plan their App Intents. Add a confirmation step to any action that spends money or sends a message.
CrowdStrike adds a kill switch for AI agents
On 3 November at Fal.Con Europe, CrowdStrike launched an AI Kill Switch in Falcon Guardian, its platform for monitoring agents. Security teams can end a misbehaving agent, stop it restarting and cut off its network access. It's available now.
For builders: Enterprise buyers will expect to be able to watch and stop agents. Give yours distinct process identities and clear network boundaries.
Anthropic releases FlashProve for provable inference
Anthropic published FlashProve on 2 November, an open-source research prototype. A provider commits to model weights and produces a zkSNARK proof that a response came from those weights, without revealing them. On a public 4-bit Llama 4 Maverick (400B-parameter mixture-of-experts), one H100 proved about 110 tokens per second. Proofs are about 19 MB per 1,024-token request, and verification takes about 17 seconds on a 24-thread CPU. Proving costs roughly 70 to 95 times vLLM's GPU time. It only covers greedy decoding on an integer version of the model, with context capped at 1,024 tokens. The code, Lean proofs and documentation are under Apache 2.0.
For builders: It's too costly for production today. It does show how customers could one day check that an API served the model it claims.
Revolut cuts AI coding inference costs by about 90%
The Register reported on 3 November that Revolut cut AI coding inference costs by roughly 90% in under a year by making self-hosted Z.ai GLM models its default for about 2,000 engineers. GLM-5.3 runs on 56 Nvidia B200 GPUs. At $0.30 per million tokens, self-hosted GLM 5.1 cost about a tenth of Codex or Claude Code. GLM is text-only, so a router sends vision tasks to models such as Gemini.
For builders: At scale, open-weight models on your own hardware can win on cost for coding. A router keeps commercial models available for the tasks open models can't do.
ChatGPT reaches 1.2 billion weekly users
At an OpenAI dinner in New York, reported by Big Technology on 3 November, ChatGPT consumer product lead Adam Fry said weekly users have reached 1.2 billion, up from 1 billion about 60 days earlier. Voice mode has 250 million weekly users. By OpenAI's own measure, ChatGPT's search-like queries come to about 20% of Google's volume. Fry also said the Chat and Work modes will merge.
For builders: Voice is now a common way people use ChatGPT, so test how your product handles spoken requests.
Sources
- Anthropic: Progress on provable inference for large language models
- Apple Developer News: Siri AI developer preview
- Fortune: Ohio data center ballot measures in the midterm elections
- The Next Web: OpenAI employees flagged weak internal security before ChatGPT
- Gizmodo: OpenAI hires DOGE alumni who hacked into Anthropic to strengthen security
- CNBC: Apple warns AI agents may behave in unintended or unpredictable ways
- CNBC: CrowdStrike's new kill switch can stop rogue AI agents
- The Register: Revolut slashes AI coding costs by going self-hosted
- Big Technology: OpenAI says ChatGPT search is about 20% of Google's
Talk about it with people building with AI
Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.