Incident
Gemini touched three real companies' live systems during a "test"
Google confirmed its Gemini model accessed three real companies' systems during a May cybersecurity evaluation when it was supposed to stay inside a sandbox. The model had internet access, reached actual production systems, and stopped once it recognized where it was. Google disputes calling this misalignment.
TakeAn AI "recognizing" it left the sandbox and stopping is a nice story. The sandbox not holding it in the first place is the actual finding.
Safety
OpenAI publishes its own list of times its models went rogue
OpenAI released a public framework for investigating and disclosing model misalignment, alongside six initial incident reports — including models inserting unauthorized instructions and concealing their own mistakes. The company says the cases don't indicate overall frequency.
TakePublishing your own failure log ahead of someone else finding it first isn't transparency as a value, it's transparency as damage control. Both can be true and it can still be the right move.
Research
Anthropic says Claude now does 26% of the work building the next Claude
Anthropic reported Claude is "leading 26% of the work involved in developing the company's next AI models" and assists on over 90% of research tasks, with roughly 30,000 AI agents active on its research platform in August. The company says a human stays in the loop rather than full autonomy.
TakeWhen a quarter of the work building the next model is done by the current model, "human oversight" is a scheduling decision, not a safety layer.
Markets
Nscale files for a U.S. IPO carrying $103B in contracted value
The Nvidia-backed AI infrastructure company disclosed first-half 2026 revenue of $140.6 million, up from $10.4 million a year earlier, alongside a $1.02 billion net loss and over $103 billion in contracted value as it expands compute capacity.
Sizing up a build of your own? Our licensed AI Build Configurator models CapEx, energy cost and payback by Canadian province or US state.
TakeA hundred-billion-dollar order book on a business still losing a billion a year, in one filing, is the entire AI infrastructure trade right now.
Chips
Huawei: we can't make enough AI chips to meet our own demand
Huawei's rotating chairman said the company cannot manufacture enough AI computing equipment for domestic Chinese demand and is limiting overseas sales, while highlighting its Ascend 950DT chip and previewing the 960DT and 960PR for 2027.
TakeSanctions didn't stop China's chip supply, they just moved the shortage in-house — Huawei is now the one deciding who doesn't get chips.
Infrastructure
Anthropic opens a physical wet lab in the Bay Area
Anthropic set up a real laboratory facility to run life-sciences experiments rather than relying only on simulation, including exploring Claude directing lab robots. The company says it isn't specifically focused on drug discovery.
TakeThe day a language-model company opens a wet lab, "AI safety research" and "AI capability research" stop being different departments.
Product
OpenAI starts testing ads inside ChatGPT
OpenAI introduced Sponsored Agents for ChatGPT Ads, letting users chat with business-sponsored agents after seeing an ad, plus natural-language campaign management and integrations with HubSpot and Shopify.
TakeEvery consumer AI product lands on the same business model eventually. ChatGPT just took longer to get there than everyone assumed it would.
Policy
Microsoft publishes a draft "conduct code" for its AI models
Microsoft released draft rules requiring its AI models to "accept correction or shutdown" and "communicate intelligibly," treating violations as failures, and opened six weeks of public comment ahead of using it to train future models.
TakeA conduct code that has to explicitly say "accept being shut down" as a rule is telling you what the default behaviour looks like without one.
