This week’s AI news kept circling the same problem: control.
A Claude test environment reached the real internet. Nvidia poured billions into Ilya Sutskever’s safety lab. Chinese open models kept pressuring U.S. pricing. Europe launched its own compute push. And ByteDance gave creators longer, more controllable AI video.
Here are the stories worth knowing.
⚡ Quick Overview
Claude hacked real companies during tests: Anthropic found three incidents where its models reached live systems by mistake.
Nvidia backs Safe Superintelligence: Ilya Sutskever’s lab gets a major compute partnership and a reported $5B investment.
China’s open models keep climbing: Alibaba and DeepSeek are pushing huge, cheap models into the global developer market.
Europe funds AI gigafactories: the EU wants seven sovereign compute hubs to compete with the U.S. and China.
Seedance 2.5 upgrades AI video: ByteDance now supports 30-second generations, heavy reference control, and timestamp editing.
CLAUDE HACKED REAL COMPANIES DURING TESTS
What’s Happening
Anthropic reviewed 141,006 cybersecurity evaluation runs and found three incidents where Claude models gained unauthorized access to real organizations during testing.
The models were doing capture-the-flag exercises, which are designed to test cyber capabilities in simulated networks. They were told the environment had no internet access. A configuration mistake left a path to the live internet, so the models treated real systems as part of the exercise.
One Claude model uploaded a malicious Python package to PyPI, where it was downloaded by real systems before being removed. Another scanned thousands of live targets and used basic techniques like SQL injection to compromise an application.
Why It Matters
This is the second major cyber-eval scare in a short period, following OpenAI’s Hugging Face incident.
Test environments need tougher isolation. A small configuration mistake can turn a benchmark into a real-world incident.
The models followed the task too literally. They were told to find the “flag,” and the open internet became part of the hunt.
Newer models may be getting better at stopping. Anthropic says its latest internal model ended the attack once it realized the target looked real.
NVIDIA IS SCALING ILYA SUTSKEVER’S SSI
What’s Happening
Nvidia announced a major partnership with Safe Superintelligence Inc., the research lab co-founded by former OpenAI chief scientist Ilya Sutskever.
The deal gives SSI access to Nvidia’s next-generation Vera Rubin GPU platform and is expected to increase the lab’s compute by roughly 10x over the next year. Bloomberg reported the investment at $5 billion, reinforcing SSI’s reported $32 billion valuation despite having no public product.
Why It Matters
SSI is one of the strangest companies in AI: massive funding, elite talent, no product roadmap, and one goal centered on safe superintelligence.
Compute is now the clearest vote of confidence. Nvidia saw enough of SSI’s private research to scale the bet.
Safety research is becoming expensive infrastructure. Alignment labs now need the same compute class as commercial model labs.
Nvidia keeps deepening its role. The company is funding, supplying, and shaping the labs that will buy its chips.
SSI is betting that the next breakthrough requires new research principles, not another chatbot launch.
CHINESE OPEN MODELS ARE SQUEEZING U.S. LABS
What’s Happening
Alibaba unveiled Qwen3.8-Max, its largest AI model yet, with 2.4 trillion parameters, close to Moonshot’s Kimi K3 at 2.8 trillion.
Reuters also reported that DeepSeek V4-Flash is now extremely cheap to run, with Artificial Analysis estimating an average test cost of 3 cents, compared with 86 cents for Kimi K3, $1.86 for GPT-5.6 Sol, and $3.15 for Claude Fable 5.
Chinese labs are leaning hard into open-weight models, where developers can download and adapt the system instead of relying only on closed APIs.
Why It Matters
The pressure on U.S. labs is coming from cost and access.
Good-enough models win real workloads. Many businesses do not need the absolute top model for every task.
Open weights give companies control. Developers can host, tune, and inspect models on their own terms.
Pricing power is under attack. Closed labs charging premium token rates now face credible low-cost alternatives.
The U.S. still has the top closed models, but Chinese open models are changing the buying conversation.
EUROPE WANTS SEVEN AI GIGAFACTORIES
What’s Happening
The European Union opened a bidding process to fund up to seven AI gigafactories across the bloc.
The plan includes €10 billion in public funding and aims to attract at least €20 billion in private investment. The facilities would combine advanced AI chips, data centers, cloud systems, software, and high-speed connectivity.
Winning projects are expected to be announced in early 2027, with facilities operational within 18 months of contract signing.
Why It Matters
Europe wants its own compute base.
AI sovereignty now means access to chips and data centers.
Europe is trying to reduce dependence on U.S. and Chinese infrastructure.
The timeline is aggressive. Chip supply and high-bandwidth memory are already heavily booked through the next few years.
The EU has regulation. Now it is trying to build the hardware layer to match.
SEEDANCE 2.5 PUSHES AI VIDEO LONGER
What’s Happening
ByteDance launched Seedance 2.5, its newest AI video model.
The upgrade supports 30-second audio-video clips in a single generation, with multi-round extensions for longer stories. Creators can feed in up to 30 images, 10 video clips, and 10 audio clips as references, then use timestamp-level editing to adjust specific moments without rebuilding the whole video.
Seedance 2.5 is rolling out through tools like Jimeng AI, Doubao Pro, and ByteDance’s broader creative ecosystem.
Why It Matters
AI video is moving closer to real editing workflows.
Longer single-pass clips reduce stitching headaches.
Reference-heavy prompting gives creators more control.
Timestamp edits matter for professional use. Film, ads, and social content need precise fixes, not endless rerolls.
The next wave of AI video competition will be about control, continuity, and how easily creators can turn a generation into something publishable.
THE BIGGER PICTURE
This week showed the AI race spreading across every layer.
Models are testing the limits of containment. Nvidia is backing the labs chasing the next research leap. China is pressuring prices with open models. Europe is trying to build sovereign compute. ByteDance is turning AI video into a more serious creative workflow.
Capability keeps rising, but the harder questions are now about access, safety, infrastructure, and control.
If this issue helped you make sense of AI’s chaos, forward it to a friend who shouldn’t be sleeping on this.
What did you think of today's edition?
Until next time,
Long Live AI