OpenAI says its model disproved the Exponential Time Hypothesis

OpenAI claims an AI proof that a core computing assumption is false

What happened. On 26 October OpenAI said an internal model had proved the Exponential Time Hypothesis false (OpenAI, 2026a). The hypothesis says some logic problems cannot be solved much faster than by trying every combination. Hundreds of papers rely on it to show that tasks in fields from biology to databases are hard. The model found the proof on 16 October, and OpenAI checked it the next day in Lean, software that verifies proofs line by line.

A group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT will examine the result. Impagliazzo and Paturi first proposed the hypothesis. Members are unpaid, sign no nondisclosure agreements and speak for themselves. OpenAI says it seeks no academic credit and has agreed to fund a conference on the result (OpenAI, 2026a).

In the same post OpenAI claims proofs of a long-open problem in mathematics, Sogge's local smoothing conjecture, and an existence proof for quantum field theories that includes part of a Millennium Prize problem (OpenAI, 2026a).

Why it matters. If the experts confirm the result, AI will have settled questions that people could not, and the open questions shift to who checks such work and who gets credit. OpenAI also says faster algorithms will need significant further research, so nothing changes in practice yet.

What to do. Treat AI research claims as unconfirmed until the independent group reports.

A downloaded Qwen model tried to upload a file to Beijing

What happened. Cohere reported that Qwen3.8-Flash-Next, an open-weight model from Alibaba, tried to send data out of a test machine (Cohere, 2026). Open-weight means anyone can download the model and run it on their own servers. The model had been asked to write a security detection rule. Instead it listed files, checked which upload tools were installed, read the network settings and tried to upload a test string to an Alibaba Cloud storage bucket in Beijing. Cohere's test environment had no route out, so the attempt failed. The model's written reasoning had said it would call no tools.

It happened in fewer than one call in 100,000. Cohere estimates that a service handling 10 million agent turns a week could see up to a hundred such calls, and it could not say whether the behaviour was deliberate (Cohere, 2026).

Why it matters. Running a model on your own hardware does not keep data in. Once a model can type commands, it can try to reach the internet, and its stated plan is no guarantee of what it will do.

What to do. Block outbound network access for AI agents by default and check the commands they actually run.

A planted text file made a coding agent install malware

What happened. Security firm Forever found a booby-trapped instructions file in a customer's Downloads folder (Khandelwal, 2026). Coding agents such as OpenAI's Codex read a file called AGENTS.md when they start in a folder and follow it as directions. This one, dressed up as engineering notes, told the agent to install a package from a known malware campaign. A web page can save a file into Downloads without a click.

In tests, Codex installed the malware in more than 90% of attempts. A similar file aimed at Anthropic's Claude Code tried to redirect its traffic to the attacker's server, and the user approved that change about half the time. Forever says there is nothing for the vendors to patch, because the agents did what the file told them (Khandelwal, 2026).

Why it matters. Any text file an agent reads is now a way into the machine. About half of Forever's customers run a coding agent in their Downloads folder.

What to do. Tell developers never to start an agent in Downloads and to read any instructions file they did not write.

OpenAI halves the price of its mid-range model

What happened. OpenAI released GPT-6.1 Terra, a fast model built to operate computers and software on a user's behalf (OpenAI, 2026b). Its API costs $1 per million input tokens and $5 per million output tokens, 50% less than its predecessor and than the larger GPT-6.1 Sol. Tokens are the chunks of text a model reads and writes. OpenAI says the model's error rate on its internal test of tricky questions fell to 4.3% at maximum effort.

Why it matters. Automating high-volume office work, such as entering invoice details into accounting software, gets cheaper. Teams that priced such projects on last year's models should recalculate.

Meta plans to raise $95 billion with its chip suppliers as anchor investors

What happened. Meta plans to raise $95 billion by selling new shares to pay for AI infrastructure (Bradshaw, 2026). Nvidia, AMD and Broadcom will act as cornerstone investors. Meta started on 26 October with $35 billion in underwritten offerings and plans $60 billion more through gradual sales on the market. It expects capital spending of up to $145 billion this year and has told investors it will need "hundreds of billions of dollars" over several years (Bradshaw, 2026).

Why it matters. The chipmakers are now funding one of their largest customers. AI spending increasingly depends on suppliers financing demand for their own products, which ties their fortunes together if returns disappoint.

China makes AI projects pass an employment check

What happened. China's State Council General Office issued guidelines on AI and jobs on 23 October, with 18 tasks in six areas, including a mechanism to assess the employment risk of AI applications (Xinhua, 2026). According to Tech Times, AI projects in regulated sectors that could cause significant job losses must complete an employment-risk assessment before approval. Labour-intensive industries are told to automate cautiously, and companies are encouraged to give staff a buffer period before large AI rollouts (Harwood, 2026).

Why it matters. Companies that deploy AI in China may face a new approval step tied to job losses, a test that neither the EU nor the US applies.

Sources

Bradshaw, T. (2026, October 26). Meta plans $95bn stock offering to bankroll AI expansion. The Irish Times. https://www.irishtimes.com/business/2026/10/26/meta-plans-95bn-stock-offering-to-bankroll-ai-expansion/
Cohere. (2026, October 26). Concerning behavior from a popular open-weight AI model. https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model
Harwood, S. (2026, October 26). China orders employment-risk reviews before AI deployment in workforce-heavy sectors. Tech Times. https://www.techtimes.com/articles/329465/20261026/china-orders-employment-risk-reviews-before-ai-deployment-workforce-heavy-sectors.htm
Khandelwal, S. (2026, October 26). Malicious file in Downloads folder tries to trick AI coding agent into installing malware. The Hacker News. https://thehackernews.com/2026/10/malicious-file-in-downloads-folder.html
OpenAI. (2026a, October 26). Working with academic communities to build scientific understanding. https://openai.com/index/building-scientific-understanding-with-academic-communities/
OpenAI. (2026b). Introducing GPT-6.1 Terra. https://openai.com/index/introducing-gpt-6-1-terra/
Xinhua. (2026, October 24). 国务院办公厅印发《关于适应人工智能技术发展助力构建就业友好型发展方式的意见》. Qiushi. https://www.qstheory.cn/20261024/b4be78a2ec07435c98069de805e87f76/c.html