OpenAI's agents produce 948 Linux kernel exploits

OpenAI's research post "A thousand fewer paths to root" was in circulation by 5 November. It covers a 10-day run in October in which agents went through about 2.8 million lines of Linux kernel code that ordinary users can reach. That's roughly 7% of the kernel. Within 12 hours the agents had logged 2,308 candidate findings, and 1,724 of those were reproduced in virtual machines. They were then grouped into 948 distinct privilege-escalation flaws, with one candidate patch per flaw. The run used an unreleased OpenAI cyber model and cost about US$1.21 million in API credits. OpenAI calls its figure of about US$1,279 per flaw an upper bound. A July run with GPT-5.6 Sol found 166 flaws at about US$3,094 each.

By the end of the October scan, other contributors had fixed 61 of these flaws upstream, had staging patches ready for 43 more, and had 299 under public discussion. OpenAI's advice is to treat a low-privileged account on an unpatched machine as "functionally equivalent to root access for a well-resourced threat actor." It also points to Firecracker and gVisor for isolation.

For builders: If your agents run untrusted code on shared Linux hosts, assume local root is cheap for an attacker. Put each workload in a microVM or gVisor sandbox and patch kernels fast.

Meta acquires Conductor

On 4 November Meta confirmed it had bought Conductor, which Charlie Holtz founded in 2024. The product runs coding agents from labs like Anthropic and OpenAI side by side in separate workspaces, with each one working on its own part of a project. You can run it locally, in the cloud or from a phone. Terms weren't disclosed. Conductor's cloud service shuts on 20 November, and it'll start deleting users' cloud data after that. Its core features keep working until early next year.

For builders: If you keep work in Conductor's cloud, export it before 20 November and choose another tool for running agents in parallel.

OpenAI agrees to buy Pierre

OpenAI agreed on 4 November to acquire Pierre Computer Company, a San Francisco startup. Pierre makes Code.Storage, a Git infrastructure layer, plus the Trees and Diffs libraries, and Codex already uses both libraries. OpenAI plans to bring Code.Storage into Codex. Current customers can keep using it until that's done. Pierre's open-source projects stay open, and OpenAI says it'll keep contributing to them. Neither company gave a price.

For builders: If you depend on Trees or Diffs, they're staying open. Expect Codex to take on more of the repository and code review work.

Three big models arrive

Tencent launched Hy4 on 4 November. It's a 770-billion-parameter model that can read charts and documents, reason across several images, and operate command-line tools and graphical apps, including Blender and CAD software. API pricing is US$0.834 per million input tokens and US$2.501 per million output tokens, and Tencent says the weights and a system card are coming soon.

Moonshot AI's Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion parameters, 104 billion of them active. It takes text and image input, has a 1,048,576-token context window and ships under its own Kimi licence. Ricursive Intelligence's Ripple-65B Preview is a dense 65B model under Apache 2.0, built mainly for agentic coding, with the same 1M-token context. A separate NVFP4 checkpoint lets it run on local hardware such as NVIDIA's DGX Spark.

For builders: Ripple's Apache 2.0 licence and local checkpoint make it a candidate for private coding agents. Read Kimi's licence before using it commercially, and run your own evals rather than relying on vendor benchmark scores.

Midterms put AI money and data centres on the ballot

Voters in at least 38 US jurisdictions approved data centre limits in the 3 November midterms. These ranged from bans and moratoriums to caps on electricity use. Nearly 80% backed restrictions in Joliet Township, Illinois. TechCrunch reported on 4 November that AI-industry PACs had mixed results. Leading the Future, backed by Greg Brockman and Andreessen Horowitz, spent US$27.38 million in the primaries, and several candidates its affiliates backed lost. Anthropic-backed Public First did better with Democrats.

On 5 November OpenAI released a National Community Compact for the campuses it owns and operates. It commits to paying the full cost of powering them, publishing site-level energy and water figures, and limiting most NDAs with officials. The compact doesn't include penalties or independent verification.

For builders: Local votes now limit where compute can be built, and that could feed into capacity and pricing. Keep an eye on the fight over whether federal law overrides state AI rules.

Claude helps solve a bacterial vault structure

On 4 November Anthropic described how Claude helped its team determine the first 3D structure of a vault particle from prokaryotes. The particle, from a bacteriophage, is built from 96 copies of one protein and holds about 25% more volume than mammalian vaults. Claude searched about 15 billion predicted proteins and handled much of the lab and cryo-EM work, while the researchers made the analysis decisions. A new purification method took about 90 minutes, down from about three days. Anthropic also listed the mistakes. One protocol error cost two weeks, and the researchers told Claude to stop improvising.

For builders: For agentic lab or data pipelines, put a human checkpoint at each analysis decision and keep a log of errors.

Canva restructures 22% of staff

On 4 November Canva said a restructuring affects 1,271 people, about 22% of its workforce, through job cuts, changed roles or relocation. CEO Melanie Perkins said growth had slowed for the first time in a decade. Earlier this year Canva lowered its expected revenue growth from 30% to about 20%.

For builders: Generative tools are putting pressure on established design software, so expect more incumbents to merge teams and change their product plans.

Sources

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.

Become a member

Get the daily AI news by RSS