Daily AI digest, November 4

Felix Laurent 5 Nov 2026 5 min read

news

The biggest AI news of November 4 came from security labs and ballot boxes. OpenAI said its agents found 948 privilege-escalation flaws in the Linux kernel in one run. Voters sided against data centers in 40 of 41 local ballot measures on election day. Meta and OpenAI each bought a startup that builds tools for coding agents. Below are the big stories of Wednesday, November 4, plus the late Tuesday news and the day's rumours.

OpenAI's agents find 948 Linux kernel flaws

OpenAI described a pipeline of AI agents that audited the Linux kernel in mid-October with an upcoming cyber model. It picked 5,034 audit targets, about 7% of the kernel, and produced 2,308 candidate findings within 12 hours. Within 30 hours, 1,724 of them were reproduced, Unite.AI reports.

The agents then wrote 1,169 working exploits, which variant analysis grouped into 948 distinct privilege-escalation flaws. The run cost about $1.2M in API credits, roughly $1,279 per flaw. In July, a run with GPT-5.6 Sol found 166 such flaws for about $514,000, Interesting Engineering reports.

The 948 figure counts candidate patches that still have to be reviewed and merged by kernel maintainers.

Voters reject data centers in 40 of 41 local measures

Results from Tuesday's elections came in on Wednesday. Voters in 10 states decided 41 local measures on data centers, CNBC reports, citing Ballotpedia. They sided against data centers in every one but Harwood, North Dakota. Ohio alone had 21 of the measures.

In a CNBC voter poll, 56% said data centers were a crucial or major factor in their vote. OpenAI already plans to call its data centers "AI compute campuses" after the midterms; we wrote about that here.

Meta and OpenAI buy tools for coding agents

Meta bought Conductor, a startup that ran parallel AI coding agents, Axios reports. Terms were not disclosed, and Conductor's cloud service will shut down on November 20.

OpenAI agreed to buy Pierre Computer, whose Git-based Code Storage already runs inside Codex. Pierre's open-source Diffs and Trees projects stay open, and the deal still needs regulatory approval, Unite.AI reports.

Cognition launched SWE-2 Flash, a 276B-parameter model with 12B active, built on Thinking Machines' Inkling-Small. Cognition says it runs three times faster than SWE-2 at a fifth of the price. It scores 20.0 on Terminal-Bench 4.0 against 27.3 for SWE-2 and is free in Devin until November 30.

Smaller updates

Nine smaller stories covered models, chips and money.

Anthropic used Claude across a biology project that produced the first structure of a bacterial vault particle. It is built from 96 copies of one protein and encloses about 25% more volume than mammalian vaults. A purification step that usually takes three days took about 90 minutes.
Tencent took Hy4 out of preview, calling it the first Hy model with native vision and computer use. Open weights are "coming soon"; more in our article.
Pangram released Pangram Image 1.0, which it says catches 99.2% of AI-generated images at a 0.27% false-positive rate. We checked where the figures come from here.
Mistral showed 133 agents running on Mistral Large 4 that found and patched a recent Django CVE. Our read of the demo is here.
Oracle canceled a five-year, roughly $7B deal to lease about 100,000 AI chips to Tencent, the Financial Times reports. We covered the politics behind it here.
Chinese firms ordered about 6 million Nvidia RTX Pro 5500 chips but expect only about 1% of them in the next four months, Bloomberg reports. More in our article.
AMD will pay about $2.9B in cash and stock for Toronto chip startup Taalas, and $1.4B of it depends on unspecified conditions, The Globe and Mail reports.
Off-balance-sheet AI obligations at Alphabet, Microsoft, Amazon, Meta and Oracle nearly doubled in six months to about $3.2 trillion, Nikkei Asia estimates. Former New York Fed chief Bill Dudley warned about the same hidden funding on Tuesday.
Xiaomi humanoid robots sorted car parts for a full 8-hour shift at its EV factory. Xiaomi did not say how many parts they sorted or how often people stepped in, Electrek notes.

Four of the nine are about chips and the money that buys them.

Late Tuesday updates

Three stories landed after our Tuesday digest went out.

Binance published its first report on the October 29 theft of about $6.34B in crypto. The attackers' path ran through AI tools, including a tampered installer for an AI coding tool; our breakdown is here.
Sam Altman wrote on X that people should stay in control and should not cede control to superintelligence, and that this should constrain AI progress.
Ai2 released StateBench, a test of tracking changing values across long documents. Gemini 3.8 Flash, Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna drop to 13% accuracy or lower on the hardest tasks.

Two of the three are about keeping AI under human control.

Rumours

Seven more stories circulated on November 3–4, but we could not confirm them with the companies or a primary source.

Moonshot's deadline to publish the open weights of Kimi K3.1 Pro was November 4. Some outlets say the weights are now public, others that they had not landed, and reports disagree on the license.
Claude is credited with two Xen hypervisor flaws that let a malicious guest read and write other virtual machines' memory, according to secondary reports of a Xen advisory. We could not read the advisory.
Palo Alto Networks is reportedly in talks to buy Clutch Security, which tracks credentials used by AI agents, for about $500M, The Information reports.
Microsoft reportedly declined an option to more than double a Texas contract with Nscale, the Financial Times reports.
Anthropic reportedly signed a licensing deal with the Associated Press to show AP election data in Claude, Adweek reports, citing two people familiar with it.
Tenspresso, a three-month-old chip startup founded by former Google TPU engineers, is reportedly in talks at a $500M valuation.
Sam Altman reportedly drew backlash for saying that keeping DRAM in consumer PCs is "quite terrible", and then changed his X name to "RAM Altman". We could not open the original reports; our explainer on how AI took PC parts is here.

Treat all seven as unconfirmed until the companies announce them.

What the day was about

Wednesday showed how far AI has moved into security work. One OpenAI run found 948 kernel flaws for about $1.2M, and Mistral showed 133 agents fixing a Django bug. Anthropic's Claude helped solve a protein structure in 34 hours.

The pushback came from voters. In 40 of 41 local measures they sided against new data centers, the buildings that make this work possible. On November 4 the industry kept buying chips and startups, while the towns asked to host its data centers mostly said no.