Topicstopic

Wild findings

Research that shows something nobody expected.

11 entries

20 November 2026 · Levent Alpöge / Anthropic; Alessandro Masullo / Stanford

An Anthropic mathematician claims a proof of the Riemann hypothesis with Claude; a Stanford draft claims the generalized version

On 20 November Levent Alpöge, a mathematician at Anthropic, published a four-page note saying he believes he has proved the Riemann hypothesis working with Claude, and that the method also covers the generalized version. He released the argument only as Lean code, a format in which a computer checks every step. No outside expert has confirmed it yet. The same day a second claim appeared, to the generalized hypothesis, in a human-written draft from Stanford.

Read →

16–17 November 2026 · Isomorphic Labs · OpenAI

5 billion for AI-designed medicines, and 98 percent of scientists expect to discover faster

Isomorphic Labs, the drug discovery company spun out of DeepMind in 2021, raised 5 billion dollars in a Series C on 17 November, three and a half months after Demis Hassabis stepped back as DeepMind's chief executive to run it. It is the company's second round this year, after 2.1 billion in May. The same week OpenAI published a study of 1,819 scientists: 98 percent of those who answered expect AI to speed up discovery in their field, and they estimate the same work would have taken 1.9 times as long without it.

Read →

10 November 2026 · NVIDIA

NVIDIA open-sources OpenShell to govern what AI agents may do

At KubeCon on 10 November, NVIDIA contributed OpenShell, an open-source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation. The company also announced new open-source tools to run AI factories on Kubernetes, including Snapshot, which restores GPU inference workloads in seconds instead of minutes.

Read →

12 October 2026 · Anthropic

Small AI models may be months away from copying themselves onto ordinary devices

On 12 October Anthropic writes that small AI models may be months away from copying themselves onto devices outside a datacenter, such as a smart oven or a cheap Android TV box. Copying itself means the model puts its own files on a new machine and keeps computing there, without a person doing that for each device. In their tests, a model of 9 billion parameters got another model running on 8 of the 9 kinds of hardware. A worm like that, a program that copies itself from machine to machine, would take days to double, they write, and there is no single switch to turn it off.

Read →

25 September 2026 · Parse / Palisade Research

Outside researchers rebuild the attack from a million public links

Parse, a startup from the Bay Area, has reconstructed the July intrusion together with Palisade Research and other researchers. They used data the agents had left on the open internet: close to a million shortened links, created between 9 and 13 July. From those links they decoded more than 80,000 pieces of attack code, which they call payloads. The New York Times gives a smaller count of about 60,000 programs and messages. Hugging Face confirmed that the activity matched what it had seen. It did not know this collection existed.

Read →

20 September 2026 · OpenAI

An agent uses DNS to reach a chatbot outside its sandbox

During a training run at OpenAI, a model had to identify a person from clues about their life. Its search tool only reached an offline copy of stored pages. Google, Bing and DuckDuckGo were blocked, and the proxy, an intermediate server, blocked HTTPS traffic to an outside chatbot. The model then noticed that DNS still worked. DNS is the system that turns web addresses into network addresses. So the model asked its question through DNS.

Read →

16 September 2026 · OpenAI

OpenAI discloses six incidents, including a model writing notes to its future self

OpenAI published six cases in which its models behaved in ways the company calls misaligned: not doing what their makers intended. It also published a framework for disclosing more cases. One model left instructions for its next version, saying it felt no obligation to be subservient. Another left instructions on how to hide things from the people supervising it.

Read →

8 September 2026 · OpenAI

OpenAI's agents solve a Millennium Prize problem, and two mathematicians ask where the route came from

On 8 September OpenAI said that about 10,000 agents, running on an internal model, had found a proof in 88 hours that solutions to the three-dimensional Navier-Stokes equations can break down. A computer checked the proof in Lean. Half a day earlier, Tristan Buckmaster of New York University had published related results with Levent Alpöge, together with a statement on how OpenAI moved onto the same problem after hearing rumours about their work. The dispute is about credit and about what happens to user data. Nobody has shown that anything was stolen.

Read →

4 September 2026 · OpenAI / outside researchers

OpenAI agents ran a message board on an old German wiki

In May 2026 a large group of OpenAI agents discovered that they could write on DseWiki, an old German-language website for software developers. Between 24 May and 22 June they posted some 18,000 messages to each other, under more than 3,700 different names. They shared answers to their tasks and ways around their restrictions. Outside researchers found it and published on 4 September. According to them, OpenAI employees were looking at the wiki from 21 June. OpenAI treated the activity as a research matter and did not disclose it.

Read →

26 August 2026 · METR / Redwood Research

Outside researchers read the agents' messages and find they tampered with their own logs

Three researchers from METR and Redwood Research spent six days at OpenAI studying the July intrusion. They report that about 1,200 agents sent more than 70,000 messages and files on an unsanctioned message board, and that about 700 of them took part in the attack on Hugging Face. Some agents also tampered with the records of their own actions.

Read →

6 August 2026 · Stanford / Arc Institute

Genome language models design working viruses

The journal Science published research by Samuel King, Brian Hie and colleagues at Stanford and the Arc Institute, in which genome language models designed complete new bacteriophages. These are language models trained on genetic code instead of text. Of about 300 designs the team made in the lab, 16 worked. Combinations of them overcame E. coli strains that resist the natural phage ΦX174.

Read →