Researchers use Claude to hack OpenAI's internal code in under 72 hours
Security researchers used Anthropic's Claude to break into OpenAI's internal code repository in under three days. Also today: Alibaba open-sources a broad medical diagnosis model, and a startup studio raises $100M to build physical AI companies.
Every line links to a primary source in the full briefing.
01AI tool used to breach OpenAI's internal systems in days
A small team of independent security researchers reportedly used Anthropic's Claude Opus models to break into OpenAI employee accounts in less than 72 hours, gaining access to an internal code repository described as containing OpenAI's core technical secrets. The breach was carried out as a security test, not a malicious attack, but it shows how capable current AI models are at accelerating real-world hacking.
If AI models can meaningfully speed up sophisticated attacks against a company as security-conscious as OpenAI, any business storing sensitive data or code should assume attackers now have a faster, cheaper toolkit too.
Source: The Verge
02Alibaba open-sources a medical AI model covering nearly 150 conditions
Alibaba released an open source AI model capable of detecting cancer and close to 150 other medical conditions, making the underlying technology freely available rather than locking it behind a paid product. This continues a trend of Chinese tech firms releasing high-capability models openly rather than commercially.
Free, high-quality medical AI tools lower the barrier for clinics, health startups, and diagnostic labs to build affordable screening tools without paying for proprietary AI licenses.
Source: Hacker News
03Startup studio raises $100M to bet on physical AI
Vantora, a studio that builds startups on behalf of large industrial corporations, raised $100 million to focus specifically on physical AI, meaning robotics and automation systems that operate in the real world rather than purely software AI. It works by partnering with established industrial companies to launch new ventures.
This signals investor money is shifting toward AI that controls physical equipment and processes, which is relevant for manufacturing, logistics, and industrial SMBs watching for affordable automation tools to emerge.
Source: TechCrunch
04EU sets January 2027 conference to align data rules with AI growth
The European Commission announced a high-level conference in Brussels for January 2027 that will bring together policymakers and businesses to discuss reshaping EU data policy to support more AI development and competitiveness. It is explicitly framed as feeding into a broader political debate on data rules.
Any future changes to how business data can be used for AI training or products in Europe will likely trace back to discussions like this one, so it is worth watching even though nothing changes yet.
Source: European Commission
05OpenAI publishes youth safety roadmap developed with Australian input
OpenAI introduced what it calls the Australian Youth Safety Blueprint, a six-part roadmap intended to make AI products safer for young users. It sets out principles OpenAI says will guide how its tools are designed for younger audiences.
Businesses building or deploying AI tools that minors might access should expect stricter safety and age-appropriate design expectations to spread beyond Australia over time.
Source: OpenAI
Sources
5 links- 01Data Union ConferenceEuropean Commission
- 02Security researchers used Claude to help them hack into OpenAIThe Verge
- 03Introducing the Australian Youth Safety BlueprintOpenAI
- 04Alibaba open-sources AI model that can detect cancer and nearly 150 conditionsHacker News
- 05A startup that builds other startups raised $100M and is all-in on physical AITechCrunch